Session 1A: Anonymous Routing and Censorship CCS '20, November 9–13, 2020, Virtual Event, USA

Censored Planet: An Internet-wide, Longitudinal

Censorship Observatory
Ram Sundara Raman, Prerana Shenoy, Katharina Kohls∗ , Roya Ensafi
University of Michigan, ∗ Ruhr University Bochum
{ramaks, pbshenoy, ensafi}@umich.edu,∗ [email protected]

ABSTRACT 1 INTRODUCTION
Remote censorship measurement techniques offer capabilities for The Internet Freedom community’s understanding of the current
monitoring Internet reachability around the world. However, op- state and global scope of censorship remains limited: most work
erating these techniques continuously is labor-intensive and re- has focused on the practices of particular countries, or on the reach-
quires specialized knowledge and synchronization, leading to lim- ability of limited sets of online services from a small number of
ited adoption. In this paper, we introduce Censored Planet, an online volunteers. Creating a global, data-driven view of censorship is a
censorship measurement platform that collects and analyzes mea- challenging proposition, since practices are intentionally opaque,
surements from ongoing deployments of four remote measurement censorship mechanisms may vary, and there are numerous loca-
techniques (Augur, Satellite/Iris, Quack, and Hyperquack). Cen- tions where disruptions can occur. Moreover, the behavior of the
sored Planet adopts a modular design that supports synchronized network can vary depending on who is requesting content from
baseline measurements on six Internet protocols as well as cus- which location.
tomized measurements that target specific countries and websites. Established efforts to measure censorship globally utilize dis-
Censored Planet has already collected and published more than tributed deployments or volunteer networks of end-user devices [7,
21.8 billion data points of longitudinal network observations over 104]. These offer direct access to some networks and can be used to
20 months of operation. Censored Planet complements existing conduct detailed experiments from those locations, but because of
censorship measurement platforms such as OONI and ICLab by the need to recruit volunteers (and keep them safe) or the minuscule
offering increased scale, coverage, and continuity. We introduce a number of accessible endpoints in many regions of interest, they
new representative censorship metric and show how time series suffer from three key challenges: scale, coverage, and continuity.
analysis can be applied to Censored Planet’s longitudinal mea- Consequently, the resulting data tends to be sparse and ill-suited
surements to detect 15 prominent censorship events, two-thirds of for discovering events and trends among countries or across time.
which have not been reported previously. Using trend analysis, we Recent work has introduced an entirely different approach that
find increasing censorship activity in more than 100 countries, and offers a safer and more scalable means of measuring global cen-
we identify 11 categories of websites facing increasing censorship, sorship. This family of measurement techniques, including Augur,
including provocative attire, human rights issues, and news media. Quack, Satellite, Iris, and Hyperquack, use network side-channels
We hope that the continued publication of Censored Planet data to efficiently and remotely detect network anomalies from tens of
helps counter the proliferation of growing restrictions to online thousands of vantage points without relying on dedicated probing
freedom. infrastructure in the field [77, 78, 93, 100, 106]. Despite overcoming
the traditional limitations of vantage point and participant selection
CCS CONCEPTS and providing an unprecedented breadth of coverage, these tech-
• General and reference → Measurement; • Social and pro- niques have some shortcomings. Each technique only focuses on
fessional topics → Technology and censorship. one particular type of blocking, and hence does not provide a com-
plete view of global censorship. Thus far, the techniques have only
KEYWORDS been evaluated on measurements conducted over a limited period of
time, and hence did not grapple with the complexities of continuous,
Empirical Security, Measurement, Censorship, Availability
longitudinal data collection and analysis. None of the techniques
ACM Reference Format: are designed to differentiate between localized censorship by a van-
Ram Sundara Raman, Prerana Shenoy, Katharina Kohls, Roya Ensafi. 2020. tage point operator and ISP- or country-wide censorship policies.
Censored Planet: An Internet-wide, Longitudinal Censorship Observatory. Moreover, they do not have mechanisms to verify censorship and
In Proceedings of the 2020 ACM SIGSAC Conference on Computer and Commu- hence may suffer from false positives.
nications Security (CCS ’20), November 9–13, 2020, Virtual Event, USA. ACM,
To overcome these challenges, we introduce Censored Planet,
New York, NY, USA, 18 pages. https://doi.org/10.1145/3372297.3417883
a global and longitudinal censorship measurement platform that
collects censorship data using multiple remote measurement tech-
niques and analyzes the data to create a more complete view of
global censorship. Censored Planet’s modular design synchronizes
This work is licensed under a Creative Commons Attribution International 4.0 License. vantage point and test list selection processes, and schedules cen-
CCS ’20, November 9–13, 2020, Virtual Event, USA sorship measurements on six Internet protocols. Censored Planet
© 2020 Copyright held by the owner/author(s). captures a continuous baseline of reachability data for 2,000 do-
ACM ISBN 978-1-4503-7089-9/20/11.
https://doi.org/10.1145/3372297.3417883 mains and IP addresses each week from more than 95,000 vantage

49
Session 1A: Anonymous Routing and Censorship CCS '20, November 9–13, 2020, Virtual Event, USA

points in 221 countries and territories, selected for their geographi- Censored Planet data will allow researchers to continuously mon-
cal diversity and the safety of remote operators. In addition, Cen- itor the deployment of network interference technologies, track
sored Planet’s design offers rapid focus capabilities that allow us policy changes in censoring nations, and better understand the
to quickly and agilely conduct more intensive measurements of targets of interference. Ultimately, we hope that making opaque
particular countries or content in response to world events. We censorship practices more transparent at a global scale counters
make data from Censored Planet available to the public in the form the proliferation of these growing restrictions to online freedom.
of up-to-date snapshots and historical data sets1 .
Since its launch in August 2018, Censored Planet has collected 2 BACKGROUND
and published more than 21.8 billion data points of baseline longi- Two decades of research on Internet censorship has illustrated it to
tudinal network observations. Complementing previous work such be both pervasive and diverse across methods, targets, regions, and
as OONI (web connectivity tests) and ICLab, Censored Planet offers timing.
widespread coverage by running measurements in 66 (42%)–173
(360%) more countries with a median increase of 4–7 Autonomous Censorship Methods. The most commonly used censorship
Systems (AS) per country. The platform’s rapid focus capability methods are shutdowns, DNS manipulation, IP-based blocking, and
has helped provide insights into important events such as the re- HTTP-layer interference. In case of Internet shutdowns, the censor
cent large-scale HTTPS interception in Kazakhstan that has helped restricts access to the Internet completely (not to a specific web-
inform policy changes by two major web browsers [64, 98, 99]. site) [31, 112]. DNS manipulation describes cases where the user
Censored Planet processes censorship measurement data to en- receives incorrect DNS replies. These can include non-routable IP
hance detection accuracy by removing false positives using cluster- addresses, the address of a censor-controlled server hosting a block-
ing techniques [100] and obtains a novel representative measure for page, or no reply at all [8]. IP or TCP layer disruption occurs when
censorship within a country through smoothing using an optimiza- network-level connections to specific destination IPs or IP:Port
tion model. We introduce techniques for analyzing the observatory tuples are dropped or reset. This method has been specifically used
data by modeling it as a time series and applying a Bitmap-based to block circumvention proxies, and is how China prevents access
anomaly detection technique for finding censorship events. Addi- to the Tor network [5]. In HTTP(S) blocking, web traffic is dis-
tionally, we use the Mann-Kendall test for detecting trends over rupted when specific keywords, like a domain, are observed in the
time. We show how these techniques, when applied on our longitu- application payload. When detected, censoring systems may drop
dinal measurements, enable Censored Planet to detect 15 prominent the traffic, reset the connection, or show a blockpage [32, 57, 100].
censorship events during its 20-month period of measurement, two- When HTTP traffic is sent over a TLS encrypted channel, the re-
thirds of which have not been reported previously. Investigation quested domain continues to be sent in the initial unencrypted
into public OONI and ICLab data further reveals that the limitations message, providing a selector for censorship (i.e. the SNI extension
of traditional volunteer-based measurement (sparse data due to low of a valid TLS ClientHello message).
continuity and limited scale) result in the absence of data related To understand the true scale and nuanced evolution of Internet
to most events detected by Censored Planet. These events reveal censorship and how it affects global Internet communication, mul-
heightened censorship in many countries, including some (such as tiple projects have built platforms to continuously collect measure-
Japan and Norway) that have previously been regarded as having ment data. The Open Observatory of Network Inference (OONI) [43,
strong Internet freedom [46]. Using trend analysis, we find increas- 104] collects measurements from end users who download, update,
ing censorship activity in more than 100 countries, particularly and run client software. The ICLab [7, 51] project uses a set of VPN
using DNS and HTTPS blocking methods. We also find 11 cate- providers to probe from a diverse set of networks. These platforms
gories of websites that are being censored increasingly, including benefit from direct access to vantage points in residential networks
provocative attire, human rights issues, and news media. and the ability to customize measurements, and they have proven
Censored Planet’s contribution is not limited to public longitudi- invaluable in measuring censorship. However, they are challenging
nal measurement data and analysis techniques; we have been using to scale, have coverage and continuity limitations, and the data
Censored Planet’s rapid focus capabilities to accommodate requests they collect tends to be sparse and unsuitable for discovering finer
for measurements from the censorship community and investigate censorship trends among countries or across time. Moreover, main-
important events in detail. In this paper, we highlight an instance of taining a distributed network involves pushing updates and new
the use of rapid focus measurement into investigating the sudden measurements to all vantage points or volunteers which may lead
blocking of Cloudflare IPs by Turkmenistan. to delays in detection of new types of censorship.
Our results demonstrate Censored Planet’s ability to create a In recent years, remote measurement techniques have shown
more complete picture of global censorship that is complementary that it is possible to leverage side channels in existing Internet pro-
to both existing platforms such as OONI and ICLab [7], as well as tocols for interacting with remote systems, and inferring whether
qualitative reports, such as the annual Freedom on the Net Report the connection is disrupted from their responses.
by Freedom House [46]. We show through data-driven analysis Remote Detection of TCP/IP Blocking. Spooky scan em-
that qualitative reports often cover only a small number of coun- ployed a side channel for determining the state of TCP/IP reacha-
tries and that there are significant increasing trends in censorship bility between two remote network hosts [37], regardless of where
in countries considered as “Free”. The continued publication of these two remote systems (e.g., site and client) are located. In the
experimental setup, the measurement machine needed to be able to
1 https://censoredplanet.org spoof packets, one of the remote hosts needed to have a single SYN

50
Session 1A: Anonymous Routing and Censorship CCS '20, November 9–13, 2020, Virtual Event, USA

backlog (i.e., no load balancers and no anycasting), and the other

remote host needed to have a single, shared, incrementing counter
for generating IP identifier values. By monitoring the progression
of this counter over time, while attempting to perturb it from other
locations on the Internet, the method detects when connections
succeed between pairs of Internet endpoints.
The technique was extended by Augur [77], demonstrating how
this channel can be used for broad, continuous blocking detec-
tion. Augur adds a host selection subsystem to ensure that it per-
forms measurements from Internet infrastructure, only considering
routers located two or more traceroute hops upstream from end
hosts and follows the ethical guidelines set out in the Menlo and Bel-
mont reports [34, 68]. Augur also makes use of statistical hypothesis
testing to limit false detection when run at scale.
Remote Detection of DNS Manipulation. There have been
many studies that explored DNS manipulation using open DNS
resolvers, most notably Satellite and Iris [78, 93]. Satellite scans
for IPv4 resolvers that have been active for more than a month,
uses clustering techniques to detect CDN deployments, and detects
incorrect DNS responses from this information. Iris is a scalable and
ethical system that identifies DNS manipulation which restricts user
access to content (not just natural inconsistencies). To achieve high Figure 1: Censored Planet Design.
detection accuracy, Iris performs both test measurements to open
DNS resolvers and control measurements to trusted resolvers and
compares the responses using several heuristics including matching
the resolved IP, HTTP content hashes, TLS certificates, AS number
and AS name, and checking whether the TLS certificate is browser- connection timeout that does not match the web server’s typical
trusted. Iris has a higher standard for minimizing risk to operators response. Hyperquack selects infrastructural servers operated by
of DNS resolvers by only choosing name servers using their DNS ISPs as vantage points using data from PeeringDB [79].
PTR records. Our adopted technique is a synthesis of Satellite and To continuously monitor censorship and accurately derive in-
Iris, built on Satellite’s engineering efforts. For simplicity, instead sights using these complex remote measurement techniques, we
of Satellite/Iris, we just use “Satellite”. need a new scalable, efficient and extensible platform. In this paper,
we introduce Censored Planet, a global and longitudinal censorship
Remote Detection of HTTP(S) Blocking. Quack uses servers measurement platform that collects censorship data using multiple
that support the TCP Echo protocol (open port 7) as vantage points remote measurement techniques and analyzes the data to create a
to detect application-layer blocking triggered on HTTP and TLS more complete view of global censorship.
headers [106]. Quack detects interference based on whether the
server successfully echoes back (over several trials) a packet con-
taining a sensitive keyword. Quack uses control measurements 3 CENSORED PLANET DESIGN
both before and after test measurements to ensure that interference To succeed as a global, longitudinal censorship measurement plat-
is caused by the keyword tested, and not due to the inconsistencies form and perform synchronized measurements on 6 different Inter-
of the network. Quack also uses Echo’s sibling Discard protocol to net protocols (IP, DNS, HTTP, HTTPS, Echo and Discard) amidst
learn the directionality of interference. Quack makes use of more the volatility and spatiotemporal variability of Internet censorship
than 50,000 available echo servers in different countries and follows and the risk associated with measuring it, Censored Planet should
ethical norms by running Nmap OS-detection scans and selecting be: scalable, continuous, synchronized, sound, extensible, and eth-
only infrastructural Echo servers in restrictive countries [46]. ical. Censored Planet must scale to cover many vantage points,
Hyperquack extends Quack by measuring HTTP and HTTPS as we know that censorship changes across countries and even
blocking on port 80 and port 443 in a scalable, longitudinal, and within regions [2, 9, 19, 85, 118]. Censorship also changes across
safe way [100]. Hyperquack detects interference on HTTP(S) traffic time, so Censored Planet must be able to run repeated measure-
by making use of publicly accessible web servers with consistent ments regularly to capture censorship events and observe changes
behavior as vantage points. Hyperquack first builds a template of a quickly [7, 38, 100]. Censored Planet must synchronize input lists
public web server’s typical response by requesting bogus domains and measurements between different measurement techniques in
that are not hosted on the server. It then sends requests with the order to achieve completeness and comparability. Censored Planet’s
HTTP "Host" header or TLS SNI extension set to a domain of inter- measurement and analysis methods should aim to avoid false pos-
est. If there is a censor blocking the domain on the path between itives and obtain an accurate representation of censorship [100].
the measurement machine and the public web server, the measure- Finally, Censored Planet’s design and measurements must satisfy
ment machine will receive a TCP RST, a blockpage, or experience a the ethical principles that we explain further in §3.1.

51
Session 1A: Anonymous Routing and Censorship CCS '20, November 9–13, 2020, Virtual Event, USA

With these design goals in mind, we opt for a modular design for In the design and implementation of Censored Planet, we care-
Censored Planet that aids in collecting and analyzing large-scale fully followed the risk-minimization practices proposed in the stud-
measurements (cf. Figure 1): ies that introduced each remote measurement technique. Chief
• Test Requests. First, scan configurations are set based on among these is the use of hosts in Internet infrastructure (e.g.,
requests from the community (e.g. customized list of domains routers two traceroute hops away from the end user (Augur), name-
from journalists for rapid focus testing) or triggers from server resolvers (Iris), infrastructural echo servers (Quack), infras-
previous Censored Planet scans in response to anomalous tructural web servers (Hyperquack)) rather than typical edge hosts,
event alerts. with the rationale that in the “unlikely case that authorities decided
• Input Scanner. We implement an input-selection subsys- to track down these hosts, it would be obvious that users were
tem that chooses a list of domains to test, a list of vantage not running browsers on them” [106], and “because these adminis-
points, and other inputs required for Censored Planet’s oper- trators are likely to have more skills and resources to understand
ation. We build this module to be flexible enough to produce the traffic sent to their servers, the risk posed to them by these
input for both longitudinal, continuous measurements, and methods is lower than the risk posed to end users” [100]. Although
for directed, exploratory measurements (§4.1). this restriction significantly reduces the pool of hosts, there are still
• Interference Scanner. This module is the core of Censored adequately many to achieve broad global coverage.
Planet’s remote measurements. It performs and monitors Additionally, we are careful to minimize the burden on remote
Internet-wide scans for detecting the interference of test hosts by limiting the rate at which we conduct measurements.
domains, ensuring scale and coverage (§4.2). For Internet-wide scans, we follow the ethical scanning guidelines
• Data Pre-processing. To ensure accuracy, we remove false developed by the ZMap [36]. We closely coordinate with our net-
positives from Censored Planet data, utilizing recently intro- work administrators and our upstream ISP. All our machines have
duced clustering techniques [100] (§5.1). WHOIS records and a web page served from port 80 that indicates
• Censorship Analysis. Since censorship policies can vary that measurements are part of a censorship research project and
within countries and regions, we build an optimization model offer the option to opt-out. Over the past 20 months of performing
for Censored Planet data that smooths diverging country- measurements, we received an average of one abuse complaint
level results and obtains a representative metric for censor- per month, some of them being automated responses generated by
ship in a country (§5.2). network monitoring tools. So far, no complaints indicated that our
• Time Series Analysis. We analyze the longitudinal data probes caused technical or legal problems, and one ISP administra-
collected by Censored Planet to automatically detect censor- tor even helped us diagnose a problem by providing a detailed view
ship events and trends (§5.3). of what they observed.
The modular design allows easy additions to Censored Planet, such
as adding new measurement techniques or performing new kinds 4 DATA COLLECTION
of analysis, an essential component of a longitudinal measurement Once we receive test requests with scan configurations, our Input
platform. Moreover, some components act as a feedback loop to Scanner and Interference Scanner perform the tasks for measure-
others; for instance, the results from our data processing module ment data collection.
inform the vantage point selection for the next round. Before ex-
plaining each of the components of our modular design in detail, we 4.1 Input Scanner
provide an elaborate discussion on the ethics of our measurements.
Our modularized design allows custom inputs for both longitudinal
measurements and more focused custom measurements based on
the configuration. The Input Scanner performs the crucial role of
3.1 Ethics synchronizing test lists across measurement techniques, ensuring
continuity in vantage points, and updating important dependencies.
Most censorship measurement studies involve prompting hosts
in censored countries to transmit data to trigger the censor. This 4.1.1 Vantage Point Selection. The Input Scanner follows the rig-
carries at least a hypothetical risk that local authorities might re- orous ethical standards introduced in §3.1 to select infrastructural
taliate against the operators of some hosts. The measurement re- vantage points for each measurement technique:
search community has considered these risks at length at many • Augur. Infrastructural routers which are two ICMP hops
workshops, panel discussions, and program committee meetings [1, away from the end-user and have a sequentially increment-
29, 56, 66, 76, 119]. Part of the outcome of these discussions is an ing IP ID value (from CAIDA ARK data [22]).
emerging consensus that remote measurement techniques can be • Satellite. Open DNS resolvers which are name servers (from
applied ethically if there are suitable protections in place, includ- Internet-wide scans).
ing technical practices to minimize risk to individuals, as well as • Quack. Infrastructural servers with TCP port 7 (Echo) or
thoughtful application of the principles in the Belmont and Menlo Port 9 (Discard) open (from Internet-wide scans).
reports [34, 68]. This community-driven approach has been neces- • Hyperquack. Web servers that have valid EV (Extended
sary in part because institutional review boards (including at our Validation) certificates (from Censys [35]).
institution) typically consider network measurement studies to be The Input Scanner applies several additional constraints to ensure
outside of their purview when they do not involve human subjects the quality of vantage points. For example, Augur only uses routers
or their personally identifiable data. whose IP ID increment is less than five to reduce noise.

52
Session 1A: Anonymous Routing and Censorship CCS '20, November 9–13, 2020, Virtual Event, USA

For our longitudinal measurements, the Input Scanner updates scheduler maintains a global vantage point work state and manages
the list of vantage points every week. We find currently active synchronization of measurements so that vantage points are not
vantage points by either scanning the IPv4 address space (in case overloaded and there is no noise introduced in measurement. This
of Quack and Satellite) or obtaining the latest data from other is important since techniques like Quack use overlapping vantage
sources (for Hyperquack and Augur). For techniques in which we points for Echo and Discard measurements.
have to select a subset of available vantage points due to resource For our longitudinal measurements, the scheduler performs
constraints, we select vantage points from different countries in reachability scans twice a week for Hyperquack, Quack and Satel-
a round-robin manner, prioritizing vantage points from the “Not lite, and once a week for Augur. Note that Augur measurements
Free” and “Partly Free” countries from the 2019 Freedom on the Net were started in November 2019. While performing scans, our health
report [46]. We also try to select vantage points from different /24 monitoring submodule logs any measurement or vantage point
networks to ensure a representative distribution inside the country. errors appropriately and ensures that overall scan statistics are as
While updating the list of vantage points, the Input Scanner expected. For instance, the health monitoring ensures that there
tries to select the same vantage points as in the previous week is enough hard disk space to store measurement data. When pre-
of measurements to ensure continuity, and replaces any vantage processing our data (§5.1), we use these errors and statistics to
points that are no longer active. This is an important step as time eliminate failed measurements. We also mark vantage points fre-
series analysis of censorship data requires data collected from the quently failing control tests for removal in the Input Scanner. For
same source. This is because censorship may vary between different rapid focus measurements, the Interference Scanner performs more
vantage points inside a country, as we show in §6.3. We evaluate in-depth scans, such as increasing the number of trials in Augur, or
the continuity in vantage point selection in §6.1. For rapid focus checking for particular certificate patterns in Hyperquack.
measurements, the Input Scanner selects vantage points at higher We employ the same technique for measurements as described
scale in specific countries. For example, we selected 34 Augur van- in §2, with some improvements. We add the capacity for testing
tage points for our rapid focus study in Turkmenistan that we do reachability to custom ports (not only on Port 80) for Augur, and
not use in our longitudinal measurements (§7.3). remove the browser-trusted TLS certificate heuristic from Satellite
as we discovered this heuristic introducing some false negatives.
4.1.2 Test List Selection. The Input Scanner selects different do-
mains for testing in longitudinal measurements and rapid focus
measurements. For longitudinal measurements, we follow the test
5 DATA PROCESSING
list selection process of previous studies [7, 78, 100, 106] and select Accurately deriving observations about censorship from raw mea-
all the domains from the Citizen Lab Global Test List (CLTL) [27]. surement data involves several important steps that have often
CLTL is a curated list of websites that have either previously been been overlooked by previous studies [7, 77, 78, 106]. Our analysis
reported unavailable or are of interest from a political or human process includes the sanitization of raw data in a pre-processing
rights perspective. At the time of writing, the list has around 1,400 step, followed by a censorship and time series analysis. We demon-
domains. We complement this list by including the top domains strate in §6 how such comprehensive analysis steps are crucial to
from the Alexa list of popular domains to test for blocking of ma- deriving accurate observations.
jor services. Totally, we test 2,000 domains per week. The Input The analysis steps of Censored Planet is shown in the bottom half
Scanner updates both of these lists weekly, and performs liveness of Figure 1. In the pre-processing step, we aggregate the raw mea-
checks in order to ensure the domains are active. Synchronizing surement results to a common schema and use recently introduced
test lists among different measurement techniques is an essential clustering techniques [100] to remove false positives. This even-
step in introducing comparability between them. Note that Augur tually provides us with confirmed instances of censorship (§5.1).
only performs tests for domains from the CLTL because of time and In the next step, we apply optimized weights to vantage points
resource constraints. For rapid focus, our Input Scanner selects do- to ensure they are representative for the state of censorship in a
mains based on the specific event being investigated. For example, particular country, after which we obtain a measure of censorship
we selected many IPs of DNS-over-HTTPS services and Cloudflare per country (§5.2). Finally, we perform time series analysis to find
for our rapid focus study in Turkmenistan (§7.3). anomalies and trends (§5.3).

4.1.3 Other Inputs. Our Input Scanner also generates other inputs 5.1 Pre-Processing
for specific techniques. For instance, the scanner tests whether
5.1.1 Initial Sanitization. As an initial sanitization step, we remove
the test domains are anycasted by performing measurements from
all measurements that failed due to technical issues, such as loss of
geographically-distributed machines, as this information is required
measurement machine connectivity and file system failures using
by Augur to detect certain kinds of blocking [77]. The Input Scanner
health monitoring information from the Interference Scanner (§4.2).
also verifies that all the dependencies required by the measurement
techniques such as the ZMap blacklist [36] are up to date.
5.1.2 Aggregating to Common Schema. Censored Planet collects
synchronized censorship measurement data on six Internet proto-
4.2 Interference Scanner cols which enables unified analysis of global Internet censorship.
The Interference Scanner first ensures that our machines are ready Since each measurement technique collects different measurement
to perform measurements. This includes verifying spoofing ca- data (such as resolved IP in case of Satellite and HTML response
pability and ensuring the absence of firewalls. Our measurement in case of Hyperquack), we need to design a common aggregated

53
Session 1A: Anonymous Routing and Censorship CCS '20, November 9–13, 2020, Virtual Event, USA

schema to introduce comparability and interoperability for the re- the raw data set, and of this we mark around 1.5 billion (7 %) as
sults. We attribute all measurements performed in a week to the blocked. The false positive filtering removes around 500 million
start of the week (Sunday) and model our common schema as: measurements from the this set, which leaves around 1 billion con-
id | protocol | date | vp | domain | blocked firmed blocked measurements. After this stage we consider only
the confirmed cases of blocking as censorship.
Based on the vantage point (vp) and the domain tested, we also
collect and add metadata such as the country and the AS of the 5.2 Censorship Analysis
vantage point, and the topic category of the website hosted at the
Censorship policies and methods can vary in different networks
domain. We obtain country information from Maxmind [62] and
inside the country [85, 118], complicating the analysis process. For
combine data from Maxmind, the Routeviews project [91], and
example, ISPs in Russia use various methods and policies to en-
Censys [35] for obtaining AS information. Country information
act censorship, and thus users experience differences when they
was available for 99.96 % and AS information for 99.86 % of vantage
connect to distinct networks [85]. Organizational policies further
points. For the domains, we refer to the pre-defined categories of
exacerbate the issue, causing a wide range of blocking patterns
CLTL [27], and use the Fortiguard URL classification service [45]
in measurements from different vantage points inside the coun-
for the remaining Alexa domains. Our category information spans
try [100]. We provide a thorough evaluation of heterogeneity in
33 topics and covers 99.3 % of the test domains.
blocking within a country in §6.3. To ensure a representative mea-
5.1.3 Removing False Positives. Although we perform control mea- sure of censorship within a country, i.e., avoiding the effects of
surements for all of our techniques (§2), some benign responses may outlier vantage points subject to a harsher or more lenient policy
still get classified as censorship. For instance, Cloudflare endpoints compared to the rest of the country, we build an optimization model
frequently perform bot checks on measurements, which introduces that levels out contributions from outlier vantage points.
discrepancies between the test and control measurements. Such 5.2.1 Censorship Metric. Before performing the optimization, we
issues can affect both remote and direct measurements [100, 104]. first need to define a metric for censorship. At the lowest granularity
We use the clustering approach introduced by Sundara Raman et of an individual vantage point vp, we define the censorship in week
al. [100] to identify and filter out false positives in the measurement t as a percentage value:
results of Quack, Hyperquack, and Satellite. Specifically, we use a
# Domains blocked
two-step clustering technique to identify confirmed instances of Censvp,t = · 100 (1)
censorship (blockpages) and false positives. The iterative classifica- # Domains tested
tion step first identifies large groups of identical HTML responses. For a more focused view of the types of content that is blocked, we
The image clustering step then uses the DBSCAN algorithm [39] to drill down Censvp,t by domain categories.
cluster dynamic HTML pages. Each cluster is then labeled as either To find an initial estimate of censorship in a country cc with n
a false positive or blockpage, achieving complete coverage. vantage points, we aggregate Equation 1 as:
Í𝑛
In our dataset, we extract all the responses marked as blocked
𝑖=1 Censvp𝑖 , t
from Quack and Hyperquack data; for Satellite, we fetch the re- Censcc,t (Raw) = (2)
𝑛
solved IP for blocked responses and then fetch the webpages of the
We use Equation 2 as a raw metric for censorship in a country, and
resolved IP. We then use existing blockpage clusters from previous
it serves as the input to our optimization model.
work [100] and extend them by creating new clusters using iterative
classification and image clustering. From our data, we form 457 5.2.2 Optimization. To obtain a representative measure of censor-
new clusters of responses, out of which 308 are blockpages, and ship within a country that is not affected by anomalous vantage
149 are potential false positives. Note that we follow an extremely points, we build a numerical optimization model to derive weights
conservative approach in confirming a blockpage, and only do so for measurement points that allow to smooth the censorship re-
when there is clear evidence of blocking on the webpage (such as sults. To perform the optimization, we assign individual weights
“<title>¡Página Web Bloqueada!</title>”). We consider all cases of 𝜔 𝑗 for each autonomous system AS 𝑗 in the data set. As an AS can
TCP resets and connection timeouts as true cases of blocking, since contribute to multiple different measurements, we first gather all
they are confirmed through the control measurements. available results of AS 𝑗 in country cc, which results in a vector
This step involves manual effort in labelling each new cluster of measurements for the same AS and country at different points
as either a blockpage or a false positive. Fortunately, our synchro- in time (AScc,𝑗,t ). In the second step, we extend the vector by the
nized measurement and analysis process reduces this effort since a target values (Censcc,t (𝑅𝑎𝑤)) for each scan in cc :
blockpage or false positive instance found in one technique’s mea-
surements can avoid redundant effort in identifying it with others. AScc,𝑗,1, Censcc,1 (Raw)
­ cc,𝑗,2, Censcc,2 (Raw) ®
©AS ª
Moreover, since each cluster is manually verified, we generate high
confidence in identifying censored measurements. For avoiding
­ . ® (3)
­
­ .. ®
®
false positives in Augur data, we use hypothesis testing at high
confidence levels (𝛼 = 10−5 ) [77]. « AScc,𝑗,t, Censcc,t (Raw) ¬
Given the subset of results for a specific AS 𝑗 , we optimize a
5.1.4 Confirmed Results. In the time from August 2018–March weight factor 𝜔 𝑗 that minimizes the discrepancies between the indi-
2020, we conducted 21.8 billion measurements. After the initial vidual measurement results and the target value. The optimization
pre-processing, we remove 1.2 billion measurements (5.9 %) from relies on the assumption that the overall blocking percentage of a

54
Session 1A: Anonymous Routing and Censorship CCS '20, November 9–13, 2020, Virtual Event, USA

0.175 Smoothed and changing vantage points. Thus, we define the absolute change
Censorship metric

0.150 Raw in censorship between two weeks (t𝑎 , t𝑏 ; t𝑎 < t𝑏 ) as:

0.125
0.100
Δ(Censvp,t𝑎 −t𝑏 ) = Censvp,t𝑏 − Censvp,t𝑎 (6)
0.075
0.050 5.3.2 Anomaly Detection in Censorship Time Series. We build our
0.025 anomaly detection models based on the absolute change in censor-
0.000 ship (cf. Equation 6). Censvp,t is highly auto-correlated (Kendall’s

Feb '19
Aug '18

Sep '18

Oct '18

Nov '18

Dec '18

Jan '19

Mar '19

Apr '19

May '19
correlation coefficient 𝜏 = 0.93, 95 % confidence level) and hence,
an extremely high absolute change in censorship is a very good in-
dicator of incidents. Since we want to find anomalies at the country
Figure 2: Smoothing Effects–An example of the raw and level, we take the weighted average of Equation 6 for all vantage
smoothed censorship metrics for Discard censorship in Pakistan. points within a country cc to calculate the change in censorship:
𝑗=1 𝜔 𝑗 · Δ(Censvp 𝑗 ,t𝑎 −t𝑏 )
Í𝑛
Δ(Censcc,t𝑎 −t𝑏 (Smooth)) = Í𝑛 (7)
country at a specific scan date should be representative. Therefore, 𝑗=1 𝜔 𝑗
we apply a Nelder-Mead optimization that uses an error function Next, we test different anomaly detection techniques regarding
to derive the best fitting weight factor: their fit for censorship measurement data. Specifically, we test em-
v
u
tÍ  2 ploying speed constraints (such as the Median Average Deviation
𝑛
𝑡 =1 AScc,𝑗,t · 𝜔 𝑗 − Censcc,t (Raw)
(MAD) [97]), likelihood models [120], exponentially weighted mov-
arg min (4) ing average models [24], and bitmap-based models [109] for anom-
𝜔𝑗 𝑛 aly detection. We find that the bitmap-based detection technique
More precisely, we use the root-mean-square error as the error func- works best for our data, and we provide a comparative evaluation
tion that measures the delta between an individual result and the with other techniques in Appendix A.1.
target value and try to minimize this error by finding a weighting We follow the procedure in Wei et al. [109] and the implementa-
factor 𝜔 𝑗 that levels out the differences. tion by [59] to construct a Bitmap-like representation of our data
As an output of this step, we receive the weighting factor 𝜔 𝑗 after discretizing it [58]. The distance between two Bitmaps BA
that is specific for each AS 𝑗 in the data set. We apply this weight and BB of size 𝑛 × 𝑛 is then given by:
to all vantage points inside that AS, i.e., for each vantage point vp 𝑗 𝑛 Õ
Õ 𝑛
belonging to AS 𝑗 and associated weight 𝜔 𝑗 , we modify Equation 2 Dist(BA, BB) = ((BA𝑝,𝑞 − BB𝑝,𝑞 ) 2 ) (8)
to obtain: 𝑝=1 𝑞=1
Í𝑛 We use an alphabet size of 4, and a lead and lag window size of 2 %
𝑗=1 Censvp 𝑗 ,t · 𝜔 𝑗
Censcc,t (Smooth) = Í𝑛 (5) of the length of the time series for calculating the distance between
𝑗=1 𝜔 𝑗 two Bitmaps sliding along the time series. The distance acts as the
for a country cc with 𝑛 vantage points. anomaly score. We explore the events with the highest anomaly
We observed that the smoothing process removes effects that score in our findings (§7).
are caused by only a handful of vantage points while preserving
5.3.3 Trend Detection. Our trend analysis provides insights on the
the effects of a widespread censorship increase. Figure 2 shows an
methods and contents that are increasingly represented in censor-
example of raw (Equation 2) and smoothed (Equation 5) censor-
ship. For the trend evaluation of Censored Planet results, we use
ship metrics for Discard censorship in Pakistan, where censorship
the modified Mann-Kendall test [48, 50] that identifies linear trends
methods are heterogeneous [71]. We observe that widespread cen-
while being robust to gaps and length differences of time series.
sorship increases (such as that in November 2018) are preserved
The Mann-Kendall test uses hypothesis testing to find upward
while those caused by rogue vantage points (such as September
or downward (or either) trends (99 % significance level). Since it is
2018 and March 2019) are smoothed out. We provide further evalu-
important to consider the absolute change for trend analysis (to
ation of the smoothed censorship metric in Appendix A.2 and use
avoid effects due to changing vantage points), we use Equation 7 to
it to report all country-level results in our findings (§7).
construct the time series for trend detection. To obtain an estimate
of the magnitude of the trend, we use the Thiel-Sen regression
5.3 Time Series Analysis
estimator [94] to calculate the slope of the trend line from the start
Continuously collecting and analyzing censorship data is a big chal- of our measurements until the end.
lenge that has not been explored in previous remote measurement
work [77, 78, 106]. Censored Planet’s longitudinal data collection 6 EVALUATION
allows us to develop methods to automatically detect events and
We first evaluate the scale, coverage, and continuity of Censored
trends in 20 months of longitudinal measurements.
Planet, highlighting the advantages Censored Planet offers over
5.3.1 Change in Censorship. As a first step in the time series anal- existing state of the art censorship measurement platforms. Then,
ysis, we analyze the change in censorship over time. We consider we show why scale is important especially for obtaining a repre-
changes at the lowest granularity (vp) to avoid the effects of adding sentative measure of censorship within a country.

55
Session 1A: Anonymous Routing and Censorship CCS '20, November 9–13, 2020, Virtual Event, USA

Augur Quack Echo Hyperquack HTTPS

80
Quack Discard Hyperquack HTTP Satellite

Num. ASes per country

50000 60
Num. vantage points used

40000
40
30000
20
20000
0
10000

CP Pot.
ICLab

Satellite

Quack

Augur
(156)

(191)

(221)
OONI

(140)
(166)
(48)

(222)
HQ

CP
(175)
0
Feb '19

Feb '20
Jul '18
Aug '18
Sep '18
Oct '18
Nov '18

Jan '19

Mar '19
Dec '18

Apr '19
May '19
Jun '19
Jul '19
Aug '19

Jan '20

Mar '20
Sep '19
Oct '19
Nov '19
Dec '19

Apr '20
Figure 4: Coverage of Platforms–ICLab data is from September
2018 and OONI (web connectivity) and Censored Planet data is
Figure 3: Number of Vantage Points Over Time–The error bars from March 2020. Outliers have been removed for comparison.
indicate the number of /24 subnets in which we do not discover
vantage points from the previous scan.
6.2 Complementing other Platforms
Censored Planet extends the global coverage, continuity, and scale
6.1 Evaluation of Scale, Coverage & Continuity of censorship events, but it is also highly complementary to es-
Censored Planet achieves global coverage with more than 95,000 tablished censorship measurement platforms such as ICLab and
vantage points performing weekly scans (cf. Figure 3). Across the OONI. For instance, Censored Planet can detect a new instance
different measurement techniques, we use 50,000 to 60,000 vantage or pattern of censorship using its diverse and extensive coverage.
points for Quack, and an initial set of 10,000 PeeringDB web servers OONI data can then be used for on-the-ground confirmation as it
for Hyperquack, which we later extend to 25,000 web servers with contains precise measurements from end-users. ICLab’s ability to
EV certificates. For Satellite, we use 15,000 to 35,000 resolvers se- run flexible, powerful probes such as performing traceroutes can
lected under ethical constraints, and time and resource limitations be used to determine technical details subject to the existence of
force us to use only 500 to 1,000 vantage points for Augur. a VPN vantage point. This flexibility and power of running client
Continuity in measurement data is important for Censored Planet software is out of reach for remote measurements.
to establish a baseline that is comparable over time. To estimate To emphasize the relevance of Censored Planet’s key unique
the continuity of our measurements, we analyze the range of /24 features, we compare our data set characteristics with ICLab’s pub-
subnets in which we were not able to discover the vantage points licly available dataset and OONI’s web connectivity dataset, both
from the previous week of scans. Overall, we find a continuity of of which meet the current state of the art and are comparable to
93 %, which means we are able to select vantage points in the same Censored Planet’s dataset (cf. Figure 4, HQ: Hyperquack, CP: Cen-
network with significantly high probability. The slightly smaller sored Planet). To create comparability, we pick data for a full month
continuity of 89 % in Quack data is caused by the variance in ZMap (ICLab: 09/2018 [51], latest available data; OONI (web connectivity
scans [36]. We measure the /24 continuity between two different test data): 03/2020 [104]; Censored Planet: 03/2020). CP Potential
scans since measuring the continuity at the vantage point level can shows the availability of vantage points and ASes that could be
be biased by DHCP policies. At the other end, the AS continuity selected without resource constraints.
between scans is extremely high (99.01 %). The high continuity as-
Countries. In comparison to ICLab (41) and OONI (156), Cen-
sures that our time series analysis can reliably detect changes in
sored Planet covers 221 countries in 03/2020, which gives us the
censorship, and allow us to analyze trends over time accurately.
ability to measure censorship in countries other platforms cannot
One of the primary contributions of Censored Planet is the wide-
reach due to lack of volunteers or ethical risks. Considering the
spread coverage of vantage points and ASes in different countries.
Freedom on the Net Report 2019 [46], Censored Planet and OONI
On average, more than 80 % of countries have more than one van-
cover data from all 21 countries considered “Not Free”, whereas
tage point in each measurement technique, and around 50 % of
ICLab can only reach four countries in this critical category.
countries have more than ten vantage points. In March 2020, Cen-
sored Planet selected a median of 39 vantage points per country AS Coverage. Censored Planet achieves a median coverage
and a maximum of 29,072 vantage points (in China) with a 75th of eight ASes per country, where OONI has four, and ICLab has
percentile value of 305. There is a long tail with countries with one AS per country. In the month of comparison, OONI gathered
many Internet-connected devices such as China, South Korea and measurements from 1,915 ASes while Censored Planet achieved an
the United States having several thousands of vantage points. Con- overall coverage of 9,014 ASes. The total number of ASes covered
sidering the number of ASes per country with at least one vantage by Censored Planet can potentially go up to 13,569.
point selected, the median value is 8, the 75th percentile value is 33 Continuity. The varying granularity of data collection among
and the maximum value is 1,427 (in the United States). different platforms makes it difficult to directly compare continuity.

56
Session 1A: Anonymous Routing and Censorship CCS '20, November 9–13, 2020, Virtual Event, USA

1.0
Proportion of countries

Italy
0.8
Russia
0.6 IP
Discard
0.4 Echo
India
HTTP
0.2 Iran HTTPS
China DNS
0.0
0 2 4 6 8 10 12 14
Coefficient of variation in Cens(Raw)

Figure 5: Coefficient of Variation Across Countries–The CDF Figure 6: Sampling Vantage Points– Relative difference from the
shows the coefficient of variation in 𝐶𝑒𝑛𝑠 (𝑅𝑎𝑤) for vantage points baseline when sampling 1-4 Satellite vantage points in each country.
within a country for all the censorship methods tested by Censored Only the interquartile range is considered for best comparison.
Planet; the annotated countries show HTTPS blocking patterns.

7 FINDINGS
Using Censored Planet, we gathered more than 20 billion measure-
We report an estimate of the continuity of measurements by aggre- ments across 95,000 vantage points, covering a period of 20 months,
gating OONI and ICLab’s data to a weekly granularity to match and measured censorship on six different Internet protocols. Our
Censored Planet. In our measurements, we have a median AS con- data processing pipeline uses robust pre-processing, censorship,
tinuity of 96 % for the comparison month. In this period, ICLab and time series analysis techniques that introduce transparency to
achieves only 64 % continuity in ASes, which might be caused by an otherwise extremely opaque field. In this section, we focus on
a large number of reported outages through VPN configuration unexplored censorship phenomena beyond previous studies [7, 104]
changes [7]. Since OONI is dependent on volunteers running mea- to emphasize the value of Censored Planet’s novel capabilities such
surements, OONI data has an even lower AS continuity of 36 %. as scale and continuous repetitive measurements. We refer to the
This emphasizes the need for a continuous measurement system Appendix B.3 for a general overview of results.
like Censored Planet that collects repetitive measurements, since
volunteer-based data collection may be extremely sporadic. 7.1 Censorship Events
So far, our results demonstrate that the strengths provided by
One of the primary contributions of Censored Planet is the ability
Censored Planet’s high coverage and continuity complements the
to collect and analyze longitudinal baseline measurement data and
powerful detection capabilities of ICLab and OONI. In the next step,
automatically detect censorship events using our anomaly detec-
we further emphasize the importance of large-scale measurements
tion technique. To showcase this ability, we first collect a list of
to accurately represent censorship in a country.
important political, economic, and lifestyle changes that occurred
in different countries during our measurement period from news
6.3 The Importance of Scale and Coverage media and reports from other platforms such as OONI [104] and Ac-
Censorship policies not only vary between countries, but can also cessNow [3]. We then use the results from our time series anomaly
introduce differences within a country [78, 81, 82, 85, 118]. Conse- detection to uncover new events or extend known events.
quently, it is crucial to achieve sufficient coverage for an accurate Table 1 shows a summary of key censorship events detected by
representation of censorship inside a country. Censored Planet. The first section of the table has events that have
As a measure of variation, we calculate the coefficient of variation been reported previously at a limited scale. The second section
of Cens(Raw) (Equation 2) in the latest scan within countries with contains newly discovered events for which we were able to find a
five or more vantage points. Our results (cf. Figure 5) show that correlation with a political event. The third section contains key
some countries such as Iran and China with centralized censorship events detected using our anomaly detection technique (§5.3.2).
policies apply consistent blocking (lower left) [9, 38]. In contrast, Table 1 also includes results from a preliminary investigation into
candidates like Russia and Italy provide heterogeneous results due whether the events found by Censored Planet were present in
to a decentralized implementation of censorship [85]. Especially OONI’s public web connectivity dataset [104]. We find that most
in these heterogeneous countries, it is important to use multiple New events did not cause a censorship increase in OONI data, mostly
vantage points and smooth outliers in the results (§5.2). due to the low number of measurements (e.g. Cameroon–only 46
To underline this conclusion, we randomly sample 1–4 Satellite successful measurements collected from 2018-11-15 to 2018-12-
vantage points in each country and calculate the relative difference 15) or volunteers not running measurements continuously (e.g.
from the baseline Cens(Raw) (Equation 2). Figure 6 shows that Sudan–http://facebook.com was only tested on one day - 2019-04-
we can significantly decrease this relative difference by using a 08). This shows the value of a platform like Censored Planet that can
higher number of vantage points, hence, covering more individual run measurements repetitively and scalably to detect censorship
networks within a country. increases. We also investigated ICLab’s published data [51], but the

57
Session 1A: Anonymous Routing and Censorship CCS '20, November 9–13, 2020, Virtual Event, USA

Table 1: Key Censorship Events Detected by Censored Planet. Key: —Confirmed increase in blocking of at least one domain tested
by OONI, —Unconfirmed incease in blocking of at least one domain tested by OONI, —Unconfirmed blocking (but no clear censorship
increase) of at least one domain tested by OONI, —No presence of related blocking in OONI data.

Country Period Method Anomaly Category or Domain blocked Event Other Reports Presence
Score in OONI
Egypt 26 Sep 2019 HTTP, HTTPS 2.74 News Media Protests [14] OONI [86]
Iran Mar 2020 HTTP, Echo - wikimedia.com, wikia.com Policy [69] OONI [44]
Sri Lanka 21 Apr–12 May 2019 HTTP, HTTPS 3.29 Social Networking Terrorism [16] AccessNow, Net-
blocks [70, 101]
Venezuela 12–29 Jan 2019 HTTP, HTTPS 3.13 Social Networking, wikipedia.org Unrest [10, 103] OONI [11]
Zimbabwe 20 Jan 2019 HTTP, HTTPS 3.3 Social Networking Protests [17] OONI [117]
Ecuador 8 Oct 2019 DNS 3 Social Networking Protests [102] New
India 6 Sep 2018 DNS 3.14 Online Dating Law [54] New
Israel May 2019–Jun 2019 DNS - Foreign Relations and Military Conflict [111] New
Japan 28 Jun 2019 DNS, Echo 3.25 News Media Summit [21] New
Poland 22 Jul 2019 DNS, HTTP, HTTPS 3.2 Govt., News Media, Human Rights Unrest [90] New
Sudan 11 Apr 2019 HTTP, HTTPS 3.29 Social Networking Unrest [15] New
Cameroon 25 Nov 2018 HTTP 3.44 Gambling Unknown New
India Feb–Mar 2020 Echo, HTTPS 3.29 Illegal Unknown New
Italy 22 Dec 2019 Discard 3.44 Human Rights Unknown New
Norway Dec 2019–Mar 2020 DNS 3.45 Multiple Unknown New

2.5 censorship remained unusually high through April, and then spiked
HTTP HTTPS
2.0 again in the week of May 12, 2019. This contrasts most reports
Cens(Smooth)

1.5
claiming that the social media ban was lifted by May 1st [3, 70].
Our observations stress the importance of continuous and repetitive
1.0
longitudinal measurements.
0.5

0.0 7.1.2 Uncovering New Events: DNS Blocking in Norway. Norway

Feb '19

Mar '19

Apr '19

May '19

Jun '19

Jul '19

is ranked #1 (Most Free) in the Reporters Without Border Press

Freedom Index [88]. However, recent laws passed in the coun-
try encourage the blocking of websites featuring gambling and
pornography [23, 40], which led ISPs to start performing DNS
Figure 7: Social Networking Censorship in Sri Lanka. blocking [23, 60]. Our anomaly detection alerted us to high scores
in DNS blocking starting from December 2019 until March 2020
(cf. Table 1). We therefore analyze Satellite data during that period.
timeframe of ICLab measurements overlaps with only the first two Censored Planet data reveals extremely aggressive DNS blocking
months of Censored Planet measurements. Only one event (India–6 of many domains in Norway, with many blocks being consistent in
September 2018) in Table 1 falls under this timeframe, and we did all of our vantage points. During the four month period of increased
not find evidence of any DNS blocking in India in ICLab data during censorship, 25 ASes observed blocking of more than 10 domains in
that time. We next describe two events from Table 1 in detail. at least six categories. We observed the most rigorous activity in
AS 2116 (CATCHCOM), where more than 50 domains were blocked.
7.1.1 Extending Events: Social Media Blocking in Sri Lanka. On The large number of categories being targeted shows that ISPs
April 21, 2019, several bomb blasts targeting churches and hotels in Norway are not only restricting pornography and gambling
resulted in the death of more than 250 people in Sri Lanka [16, 55]. In websites, as previously thought. Indeed, the most blocked domains
response to these deadly attacks, the government declared a state included search engines (163.com), online dating sites (match.com),
of emergency, enforced curfews, and blocked access to popular and the website of the Human Rights Watch (hrw.org). The DNS
social media, allegedly to prevent the spread of misinformation blocking in Norway also shows a highly increasing trend from
and panic [55, 101]. NetBlocks and AccessNow found seven social the beginning of our measurements. Our observations show the
media websites including Facebook, WhatsApp, and Instagram to importance of measurements in countries previously thought as free.
be blocked [70, 101].
Censored Planet detected a large increase in HTTP(S) censorship
(from 0.1 % to 2 %) first on the week of April 21, 2019 (the day of 7.2 Trend Analysis
the attack) for social media content (cf. Figure 7). We observed In this section, we discuss some primary findings from our trend
22 domains (compared to 7 reported previously) being blocked, analysis of censorship data.
including domains like twitter.com that were not reported. Five out
of these 22 domains were only from the Alexa test list, showing 7.2.1 Trends in Methods. First, we consider the trends in censor-
that variety in test lists is important. After the initial peak, HTTPS ship methods. Our key findings are as follows (cf. Figure 8):

58
Session 1A: Anonymous Routing and Censorship CCS '20, November 9–13, 2020, Virtual Event, USA

0.15
Up Down
Turkmenistan
0.10 Portugal
Slope of trend line

Uzbekistan
Vatican City
0.05

0.00

−0.05
Comoros

−0.10 UAE
Discard Echo HTTP HTTPS DNS
(43,16) (20,24) (41,41) (61,46) (123,24)

Figure 8: Trends in Censorship Methods– X-axis parentheses Figure 9: Average Censorship in Freedom on the Net 2019
shows (number of countries with upward trend, number of coun- Categories–The boxplots show the range of values in different
tries with downward trend). Countries with no statistically signif- weeks across our measurement period. NC–Not Considered.
icant trend are excluded. DNS censorship slope for China is not
included, as the value is extremely high (0.93).
7.2.3 Freedom on the Net Report. The annual Freedom on the
Net Report provides a qualitative ranking of countries in three
• DNS censorship is heavily used in countries like Iran (Fig- categories (Free, Partly Free, Not Free) [46]. The annual reports have
ure 8 trend line slope - 0.048), China (slope - 0.93), and Russia been used by numerous studies in the past as an authoritative source
(slope - 0.003) because of the ease of blocking [8, 9, 38, 41, to select countries for measurement and to compare results [7, 78, 95,
78, 85]. Recent reports suggest the export of their censorship 100, 106]. However, the reports are qualitative and often cover only
models to more than 100 countries [108] including countries a small number of countries. The quantitative results of Censored
like Turkmenistan (slope - 0.15), and hence we observe an Planet extend the insights of this report by significantly increasing
overall increase in DNS censorship in 123 countries in total. the number of countries covered, and by providing concrete results
• HTTPS censorship also observed an increasing trend. Fully on the extent of access limitations. For example, the 2019 Freedom
encrypted traffic has been cited as the reason for decreasing on the Net report covers only 65 countries which is around 28 % of
censorship in the past [7], but new methods for blocking the countries tested by Censored Planet. We apply our trend and
fully encrypted traffic leads to an increasing number of coun- censorship analysis to the Freedom House categories (cf. Figure 9):
tries with higher blocking [100]. The country with the most • Not Free countries have the highest censorship rates, mainly
increasing HTTPS censorship is Uzbekistan (slope - 0.041). caused by the restrictive policies of Iran and China [9, 38].
• Discard measures censorship in one direction (Measure- Our results confirm the qualitative assessment.
ment machine → Vantage Point). An increase in the observed • Free countries show an upward trend in censorship. Exam-
rates indicate blocking independent from the direction of ples of this are Australia and the United Kingdom [12, 63].
measurement [100, 106]. Countries like Portugal have shown • Not Considered countries also show a non-negligible amount
a high increase in Discard censorship (slope - 0.045). of censorship and a comparatively more upward trend, which
The increasing trend in multiple censorship methods encourages di- suggests that the scale of Censored Planet can complement
verse measurements and highlights the importance of a unified plat- manually-compiled reports significantly.
form measuring censorship on multiple protocols synchronously.
7.3 Case Study: Turkmenistan
7.2.2 Trends in Domains. We analyze trends in the categories of Turkmenistan, a country that has been ramping up its censorship at
domains blocked to find whether some type of content is more an alarming rate, is ranked second-to-worst in the 2019 Reporters
increasingly blocked than others. without Borders Press Freedom Index [89], and was recently in
• News media censorship shows a surprising upward trend. the limelight for censoring media regarding the COVID-19 pan-
The countries with the highest increase in news censorship demic [72, 87]. In mid-April 2020, we received requests from a major
include Pakistan and Albania. circumvention tool to investigate suspected IP blocking of DNS-
• Benign categories such as gaming, media sharing, and host- over-HTTPS (DoH) servers used by its system in Turkmenistan.
ing and blogging platforms also experience an upward trend One of these DoH servers was operated by Cloudflare and since
in addition to sensitive topics like provocative attire and any Cloudflare IP allows users to reach its DoH service, we sus-
human rights issues. pected that all Cloudflare IP addresses were being blocked which
The increasing trend in blocking of benign categories highlights the would restrict access to a wide range of services. We used Censored
importance of repetitively testing all types of content for a compre- Planet’s rapid focus capabilities to run custom Augur measurements
hensive picture of global censorship. in Turkmenistan on April 17 2020, where we tested the reachability

59
Session 1A: Anonymous Routing and Censorship CCS '20, November 9–13, 2020, Virtual Event, USA

to 15 IPs (including the DoH services and Cloudflare IPs [28]) from level [77, 100, 106]. Finally, like all previous work [77, 78, 104, 106],
34 vantage points. Our results confirm that all tested Cloudflare IPs we use off-the-shelf geolocation databases that are known to some-
were blocked in at least 18 vantage points. We found interference times be inaccurate. We have used independent data sources to
in both directions of communication (inbound on anycasted IPs, confirm location accuracy in particularly critical case studies.
outbound on non-anycasted IPs), which primarily took place in We are aware that a sophisticated censor might attempt to block
the state-owned AS 20661 (TURKMENTELECOM-AS). This affects or evade our techniques, maybe by detecting and dropping traffic
more than 90 % of the public IP address space in Turkmenistan [35]. from our machines, or by poisoning probe responses with mislead-
In addition to the Cloudflare IPs, the DoH server hosted by Snopyta, ing data. Due to our control measurements (e.g., testing for benign
a non-profit service provider was also blocked. This rapid focus case domains, and tracking changes in each network’s behavior over
study shows the ability of Censored Planet to run custom measure- time and across multiple vantage points), Censored Planet can avoid
ments or increase scale when investigating censorship events. some of these countermeasures, but not all of them. So far, we have
no reason to believe that any country or network has engaged in
8 RELATED WORK active evasion of Censored Planet measurements in order to hide
censorship, although a few small network prefixes have blacklisted
An abundance of work in censorship has focused on exploring
our probe traffic.
censorship policies in specific countries, either using volunteers or
While Censored Planet provides a powerful platform for un-
accessible vantage points inside the country. The Great Firewall
derstanding censorship phenomena, fully leveraging the data will
of China and Iran’s censorship regime are two of the most studied
require much additional work, including collaboration with in-
censorship phenomena [8, 9, 13, 26, 30, 38, 61, 113, 116, 121]. Recent
country experts and researchers from the social and political sci-
increases in censorship in other countries have also prompted fo-
ences and other domains. Further application of methods such as
cused studies, such as in Russia [84, 85], Thailand [47], India [118]
machine learning and data visualization will undoubtedly expose
and others [6, 25, 52, 65, 81]. There has also been a long line of work
more insights from the data. All of these represent opportunities
on measuring Internet shutdowns, which have been increasing in
for future collaboration, both inside and outside computer science,
many countries [31, 53, 96, 112].
and are exciting avenues to explore. Our roadmap includes several
Censorship measurement platforms that focus on coverage in
features that we hope will facilitate such collaborations. We are
multiple countries have also been proposed. In addition to ICLab [7]
building a Censored Planet search interface and API that provides
and OONI [104], there are other platforms that have been active
interactive queries and integration with other platforms.
in the past, but few are still active and collect longitudinal data.
Encore [20] induced web clients around the world to perform cross-
origin requests when users visit certain websites, and the approach 10 CONCLUSION
has spurred a long line of discussion on the ethics of censorship In this paper, we introduced Censored Planet, a global censorship
measurement [34, 56, 76, 106]. The OpenNet Initiative (ONI) [73] observatory that overcomes the scale, coverage, and continuity lim-
published several reports on Internet censorship in different coun- itations of existing platforms. Using multiple remote measurement
tries before becoming defunct in 2011 [74, 75]. UBICA [4] and techniques, Censored Planet has collected more than 21 billion data
CensMon [95] used distributed PlanetLab nodes [80] and volun- points over 20 months of operation. We built representative metrics
teer deployments to perform censorship measurements in different and time series analysis techniques to discover 15 key censorship
countries, but have not been used longitudinally. events and analyze trends in censorship methods and censored
An important component of these censorship measurement stud- content, and we used Censored Planet’s rapid focus capabilities
ies is the test list of URLs and several studies have focused on for case studies of particular censorship events. We hope that Cen-
generating an optimal list of domains for testing [27, 92, 110]. The sored Planet can enhance Internet freedom by helping to bring
literature on censorship circumvention is also rich with work on transparency to censorship practices and supporting research, tool
both long-standing systems such as Tor [33], and newer systems us- development, and advocacy that seeks to protect the human rights
ing packet manipulation strategies [18, 83, 107], crowdsourcing [67], of Internet users around the world.
and strategies to disguise the destination [42, 49, 114, 115].
ACKNOWLEDGMENTS
9 LIMITATIONS AND FUTURE WORK We thank the shepherd Nicolas Christin and the anonymous review-
Like the remote measurement techniques on which our data is ers for their helpful feedback. Censored Planet’s operation is possi-
based, Censored Planet has a few inherent limitations. Even with ble because of the help and support of the exceptional sysadmins
our large global coverage, our vantage points are not fine-grained at University of Michigan and Michalis Kallitsis at Merit Network.
enough to measure every local instance of censorship, especially We thank Reethika Ramesh, Adrian Stoll, and Victor Ongkowijaya
those applied very close to end-users, such as in schools or work- for their contribution in building the platform, and David Fifield
places. Some of the remote measurement techniques have more and J. Alex Halderman for insightful discussions. We also thank
specific technical limitations: some cannot detect unidirectional Vinicius Fortuna, Sarah Laplante and the Jigsaw team for alerting
blocking (Hyperquack, Quack-Discard) or blocking of domains us to censorship events and help with Google cloud infrastruc-
that normally are anycasted (Augur), though we note that recent ture. Katharina Kohls was supported by DFG EXC 2092 CaSa –
studies have claimed that it is challenging for censors to block 39078197. This work was supported in part by the U.S. National
network traffic in a specific direction, especially at the national Science Foundation Award CNS-1755841.

60
Session 1A: Anonymous Routing and Censorship CCS '20, November 9–13, 2020, Virtual Event, USA

REFERENCES [33] R. Dingledine, N. Mathewson, and P. Syverson. Tor: The second-generation

[1] NS Ethics ’15: Proceedings of the 2015 ACM SIGCOMM Workshop on Ethics in onion router. Technical report, Naval Research Lab Washington DC, 2004.
Networked Systems Research, 2015. [34] D. Dittrich and E. Kenneally. The Menlo Report: Ethical principles guiding
[2] N. Aase, J. R. Crandall, A. Diaz, J. Knockel, J. O. Molinero, J. Saia, D. S. Wallach, information and communication technology research. Technical report, U.S.
and T. Zhu. Whiskey, weed, and wukan on the world wide web: On measuring Department of Homeland Security, 2012.
censors’ resources and motivations. In FOCI, 2012. [35] Z. Durumeric, D. Adrian, A. Mirian, M. Bailey, and J. A. Halderman. A search
[3] AccessNow. We defend and extend the digital rights of users at risk around the engine backed by Internet-wide scanning. In Proceedings of the 2015 ACM
world. https://www.accessnow.org/. SIGSAC Conference on Computer and Communications Security, 2015.
[4] G. Aceto, A. Botta, A. Pescapè, N. Feamster, M. F. Awan, T. Ahmad, and S. Qaisar. [36] Z. Durumeric, E. Wustrow, and J. A. Halderman. ZMap: Fast internet-wide
Monitoring Internet censorship with UBICA. In International Workshop on scanning and its security applications. In 22nd USENIX Security Symposium,
Traffic Monitoring and Analysis. Springer, 2015. 2013.
[5] S. Afroz and D. Fifield. Timeline of Tor censorship, 2007. http://www1.icsi. [37] R. Ensafi, J. Knockel, G. Alexander, and J. R. Crandall. Detecting intentional
berkeley.edu/~sadia/tor_timeline.pdf. packet drops on the Internet via TCP/IP side channels. In International Confer-
[6] Y. Akdeniz. Internet content regulation: UK government and the control of ence on Passive and Active Network Measurement, 2014.
Internet content. Computer Law & Security Review, 2001. [38] R. Ensafi, P. Winter, A. Mueen, and J. R. Crandall. Analyzing the Great Firewall
[7] A. Akhavan Niaki, S. Cho, Z. Weinberg, N. P. Hoang, A. Razaghpanah, of China over space and time. Proceedings on Privacy Enhancing Technologies,
N. Christin, and P. Gill. ICLab: A Global, Longitudinal Internet Censorship 2015.
Measurement Platform. In IEEE Symposium on Security and Privacy (SP), 2020. [39] M. Ester, H.-P. Kriegel, J. Sander, X. Xu, et al. A density-based algorithm for
[8] Anonymous. Towards a comprehensive picture of the Great Firewall’s DNS discovering clusters in large spatial databases with noise. In KDD, 1996.
censorship. In 4th USENIX Workshop on Free and Open Communications on the [40] R. Falkvinge. Norwegian politicians want to censor the Internet, because Rule 34
Internet (FOCI 14), 2014. (“because all the pornography”), 2016. https://www.privateinternetaccess.com/
[9] S. Aryan, H. Aryan, and J. A. Halderman. Internet censorship in Iran: A first blog/norwegian-politicians-want-censor-internet-rule-34-pornography/.
look. In 3rd USENIX Workshop on Free and Open Communications on the Internet [41] O. Farnan, A. Darer, and J. Wright. Poisoning the well: Exploring the great
(FOCI 13), 2013. firewall’s poisoned dns responses. In Proceedings of the 2016 ACM on Workshop
[10] A. Azpúrua. Wikipedia bloqueada en CANTV desde el 12 de Enero., 2019. on Privacy in the Electronic Society, 2016.
https://vesinfiltro.com/noticias/wikipedia_2019-01/. [42] D. Fifield, C. Lan, R. Hynes, P. Wegmann, and V. Paxson. Blocking-resistant
[11] A. Azpúrua, M. Chirinos, A. Filastò, M. Xynou, S. Basso, and K. Karan. From communication through domain fronting. Proceedings on Privacy Enhancing
the blocking of Wikipedia to Social Media: Venezuela’s Political Crisis, 2019. Technologies, 2015.
[12] D. E. Bambauer. Filtering in Oz: Australia’s foray into Internet censorship. U. [43] A. Filastò and J. Appelbaum. OONI: Open Observatory of Network Interference.
Pa. J. Int’l L., 2009. In 2nd USENIX Workshop on Free and Open Communications on the Internet (FOCI
[13] D. Bamman, B. O’Connor, and N. Smith. Censorship and deletion practices in 12), 2012.
chinese social media. First Monday, 2012. [44] A. Filastò, M. Xynou, and N. Fatemi. Iran temporarily blocks the Farsi lan-
[14] BBC. Egypt: Protests and clashes enter second day, 2019. https://www.bbc.com/ guage edition of Wikipedia, 2019. https://ooni.org/post/2020-iran-blocks-farsi-
news/world-middle-east-49786367. wikipedia/.
[15] BBC. Omar al-Bashir: Sudan military coup topples ruler after protests, 2019. [45] FortiNet. Fortiguard labs web filter. https://fortiguard.com/webfilter.
https://www.bbc.com/news/world-africa-47891470. [46] Freedom House. Freedom on the net report 2019. https://freedomhouse.org/
[16] BBC. Sri Lanka attacks: More than 200 killed as churches and hotels targeted, countries/freedom-world/scores, 2019.
2019. https://www.bbc.com/news/world-asia-48001720. [47] G. Gebhart and T. Kohno. Internet censorship in Thailand: User practices
[17] BBC. Zimbabwe protests: Crackdown is just a ’taste of things to come’, 2019. and potential threats. In IEEE European Symposium on Security and Privacy
https://www.bbc.com/news/world-africa-46938679. (EuroS&P), 2017.
[18] K. Bock, G. Hughey, X. Qiang, and D. Levin. Geneva: Evolving censorship evasion [48] K. H. Hamed and A. R. Rao. A modified mann-kendall trend test for autocorre-
strategies. In Proceedings of the 2019 ACM SIGSAC Conference on Computer and lated data. Journal of hydrology, 1998.
Communications Security, 2019. [49] A. Houmansadr, G. T. Nguyen, M. Caesar, and N. Borisov. Cirripede: Circum-
[19] S. Burnett and N. Feamster. Making sense of Internet censorship: a new frontier vention infrastructure using router redirection with plausible deniability. In
for Internet measurement, 2013. Proceedings of the 2011 ACM SIGSAC Conference on Computer and Communica-
[20] S. Burnett and N. Feamster. Encore: Lightweight measurement of web censorship tions Security, 2011.
with cross-origin requests. In ACM SIGCOMM Conference, 2015. [50] M. Hussain and I. Mahmud. pymannkendall: a python package for non para-
[21] Businesswire. Japan Welcomes World Leaders to Its First-ever G20 Summit in metric mann kendall family of trend tests. Journal of Open Source Software,
Osaka, 2019. https://www.businesswire.com/news/home/20190630005053/en/ 2019.
Japan-Welcomes-World-Leaders-First-ever-G20-Summit. [51] ICLAB. ICLAB: Internet Censorship Lab. https://iclab.org.
[22] CAIDA. Archipelago (Ark) Measurement Infrastructure. http://www.caida.org/ [52] Indonesia introduces new internet censorship system. https://www.arabnews.
projects/ark/. com/node/1218011/world.
[23] CalvinAyre. Gambling operators scoff as Norway approves DNS-blocking, [53] Internet Outage Detection and Analysis. https://ioda.caida.org/ioda/dashboard.
2018. https://calvinayre.com/2018/05/10/business/norway-approves-gambling- [54] A. Jazeera. India decriminalises gay sex in landmark verdict, 2018.
restrictions/. https://www.aljazeera.com/news/2018/09/india-decriminalises-gay-sex-
[24] K. M. Carter and W. W. Streilein. Probabilistic reasoning for streaming anomaly landmark-verdict-180906051219637.html.
detection. In 2012 IEEE Statistical Signal Processing Workshop (SSP). IEEE, 2012. [55] A. Jazeera. Sri Lanka bombings, 2019. https://www.aljazeera.com/news/2019/
[25] A. Chaabane, T. Chen, M. Cunche, E. D. Cristofaro, A. Friedman, and M. A. 04/sri-lanka-bombings-latest-updates-190421092621543.html.
Kaafar. Censorship in the wild: Analyzing Internet filtering in Syria. In Internet [56] B. Jones, R. Ensafi, N. Feamster, V. Paxson, and N. Weaver. Ethical concerns for
Measurement Conference (IMC). ACM, 2014. censorship measurement. In ACM SIGCOMM Conference, 2015.
[26] C. Chiu, C. Ip, and A. Silverman. Understanding social media in china. McKinsey [57] B. Jones, T.-W. Lee, N. Feamster, and P. Gill. Automated detection and finger-
Quarterly, 2012. printing of censorship block pages. In Internet Measurement Conference (IMC).
[27] Citizen Lab. Block test list. https://github.com/citizenlab/test-lists. ACM, 2014.
[28] Cloudflare. IP Ranges, 2019. https://www.cloudflare.com/ips/. [58] J. Lin, E. Keogh, S. Lonardi, and B. Chiu. A symbolic representation of time
[29] J. R. Crandall, M. Crete-Nishihata, and J. Knockel. Forgive us our syns: Technical series, with implications for streaming algorithms. In Proceedings of the 8th ACM
and ethical considerations for measuring internet filtering. In NS Ethics@ SIGMOD workshop on Research issues in data mining and knowledge discovery,
SIGCOMM, 2015. 2003.
[30] J. R. Crandall, D. Zinn, M. Byrd, E. T. Barr, and R. East. ConceptDoppler: A [59] LinkedIn. Luminol: Anomaly Detection and Correlation library. https://github.
weather tracker for Internet censorship. In Proceedings of the 2007 ACM SIGSAC com/linkedin/luminol.
Conference on Computer and Communications Security, 2007. [60] N. Macedo. Norway plots DNS blocking and further restrictions on payments,
[31] A. L. Dahir. Internet shutdowns are costing African governments more than we 2018. https://egr.global/intel/news/norway-plots-dns-blocking-and-further-
thought. https://qz.com/1089749/internet-shutdowns-are-increasingly-taking- restrictions-on-payments/.
a-toll-on-africas-economies/, 2017. [61] R. MacKinnon. China’s censorship 2.0: How companies censor bloggers. First
[32] J. Dalek, B. Haselton, H. Noman, A. Senft, M. Crete-Nishihata, P. Gill, and R. J. Monday, 2009.
Deibert. A method for identifying and confirming the use of URL filtering [62] MaxMind. https://www.maxmind.com/.
products for censorship. In Internet Measurement Conference (IMC). ACM, 2013. [63] T. McIntyre. Internet censorship in the united kingdom: National schemes and
european norms. Law, Policy and the Internet (Hart Publishing, 2018 Forthcoming),
2018.

61
Session 1A: Anonymous Routing and Censorship CCS '20, November 9–13, 2020, Virtual Event, USA

[64] Mozilla. Mozilla takes action to protect users in Kazakhstan. The Mozilla Blog, [96] R. Shandler. Measuring the Political and Social Implications of Government-
August 21, 2019. https://blog.mozilla.org/blog/2019/08/21/mozilla-takes-action- Initiated Cyber Shutdowns. In 8th USENIX Workshop on Free and Open Commu-
to-protect-users-in-kazakhstan/. nications on the Internet (FOCI 18), 2018.
[65] Z. Nabi. The Anatomy of Web Censorship in Pakistan. In 3rd USENIX Workshop [97] S. Song, A. Zhang, J. Wang, and P. S. Yu. Screen: Stream data cleaning un-
on Free and Open Communications on the Internet (FOCI 13), 2013. der speed constraints. In Proceedings of the 2015 ACM SIGMOD International
[66] A. Narayanan and B. Zevenbergen. No encore for Encore? Ethical questions for Conference on Management of Data, 2015.
web-based censorship measurement. Ethical Questions for Web-Based Censorship [98] R. Sundara Raman, L. Evdokimov, E. Wustrow, A. Halderman, and R. Ensafi.
Measurement (September 24, 2015), 2015. Kazakhstan’s HTTPS Interception, 2019. https://censoredplanet.org/kazakhstan.
[67] M. Nasr, H. Zolfaghar, A. Houmansadr, and A. Ghafari. Massbrowser: Unblock- [99] R. Sundara Raman, L. Evdokimov, E. Wustrow, A. Halderman, and R. Ensafi.
ing the censored web for the masses, by the masses. In Proceedings of the Network Investigating Large Scale HTTPS Interception in Kazakhstan. In Internet Mea-
and Distributed System Security Symposium, NDSS 2020, San Diego, California, surement Conference (IMC). ACM, 2020.
USA, 2020. [100] R. Sundara Raman, A. Stoll, J. Dalek, A. Sarabi, R. Ramesh, W. Scott, and R. Ensafi.
[68] National Commission for the Protection of Human Subjects of Biomedical and Measuring the Deployment of Network Censorship Filters at Global Scale. In
Behavioral Research. The Belmont Report: Ethical Principles and Guidelines for Proceedings of the Network and Distributed System Security Symposium, NDSS
the Protection of Human Subjects of Research. National Commission for the 2020, San Diego, California, USA, 2020.
Protection of Human Subjects of Biomedical and Behavioral Research, 1978. [101] B. Taye. Sri Lanka: shutting down social media to fight rumors hurts victims,
[69] N. Nazeri and C. Anderson. Citation filtered: Iran’s censorship of wikipedia. 2019. https://www.accessnow.org/sri-lanka-shutting-down-social-media-to-
CGCS Research, 2013. fight-rumors-hurts-victims/.
[70] NetBlocks. Social media blocked in Sri Lanka following church and hotel [102] The Guardian. Army deployed in Ecuador as protests descend into violence
bombings, 2019. https://netblocks.org/reports/social-media-blocked-in-sri- , 2019. https://www.theguardian.com/world/2019/oct/13/army-deployed-in-
lanka-following-church-and-hotel-bombings-XaAwlQBM. ecuador-as-protests-descend-into-violence.
[71] A. Nisar, A. Kashaf, I. A. Qazi, and Z. A. Uzmi. Incentivizing censorship mea- [103] The Guardian. Venezuela protests: thousands march as military faces call to aban-
surements via circumvention. In SIGCOMM. ACM, 2018. don Maduro , 2019. https://www.theguardian.com/world/2019/jan/23/venezuela-
[72] NPR. Turkmenistan Has Banned Use Of The Word ’Coronavirus’, protests-thousands-march-against-maduro-as-opposition-sees-chance-for-
2020. https://www.npr.org/sections/coronavirus-live-updates/2020/03/31/ change.
824611607/turkmenistan-has-banned-use-of-the-word-coronavirus. [104] The Tor Project. OONI: Open observatory of network interference. https:
[73] OpenNet Initiative. OpenNet Initiative. https://opennet.net/. //ooni.torproject.org/.
[74] OpenNet Initiative. Jordan, August 2009. https://opennet.net/research/profiles/ [105] Tor. Tor Browser’s default bridges, 2020. https://trac.torproject.org/projects/
jordan. tor/wiki/doc/TorBrowser/DefaultBridges.
[75] OpenNet Initiative. South Korea, August 2012. https://opennet.net/research/ [106] B. VanderSloot, A. McDonald, W. Scott, J. A. Halderman, and R. Ensafi. Quack:
profiles/south-korea. Scalable remote measurement of application-layer censorship. In 27th USENIX
[76] C. Partridge and M. Allman. Addressing ethical considerations in network Security Symposium, 2018.
measurement papers. In Workshop on Ethics in Networked Systems Research (NS [107] Z. Wang, S. Zhu, Y. Cao, Z. Qian, C. Song, S. V. Krishnamurthy, K. S. Chan, and
Ethics@ SIGCOMM), 2015. T. D. Braun. SYMTCP: Eluding Stateful Deep Packet Inspection with Automated
[77] P. Pearce, R. Ensafi, F. Li, N. Feamster, and V. Paxson. Augur: Internet-wide Discrepancy Discovery. In Proceedings of the Network and Distributed System
detection of connectivity disruptions. In IEEE Symposium on Security and Privacy, Security Symposium, NDSS 2020, San Diego, California, USA, 2020.
May 2017. [108] V. Weber. The worldwide web of Chinese and Russian information controls,
[78] P. Pearce, B. Jones, F. Li, R. Ensafi, N. Feamster, N. Weaver, and V. Paxson. Global 2019.
measurement of DNS manipulation. In 26th USENIX Security Symposium, 2017. [109] L. Wei, N. Kumar, V. N. Lolla, E. J. Keogh, S. Lonardi, and C. A. Ratanamahatana.
[79] PeeringDB. Peeringdb, 2018. https://www.peeringdb.com/. Assumption-free anomaly detection in time series. In SSDBM, 2005.
[80] PlanetLab. https://www.planet-lab.org/. [110] Z. Weinberg, M. Sharif, J. Szurdi, and N. Christin. Topics of controversy: An
[81] Portuguese ISPs given 40 days to comply with EU net neutrality empirical analysis of web censorship lists. Proceedings on Privacy Enhancing
rules. https://edri.org/portuguese-isps-given-40-days-to-comply-with-eu-net- Technologies, 2017.
neutrality-rules/. [111] Wikipedia. Gaza–Israel clashes (May 2019), 2019. https://en.wikipedia.org/wiki/
[82] List of websites/domains blocked by ISP’s in Portugal, 2019. https://tofran. Gaza%E2%80%93Israel_clashes_(May_2019).
github.io/PortugalWebBlocking/. [112] C. Williams. How Egypt shut down the internet. https://www.telegraph.co.
[83] Psiphon. Psiphon: Beyond Borders, 2020. https://psiphon3.com/en/index.html. uk/news/worldnews/africaandindianocean/egypt/8288163/How-Egypt-shut-
[84] R. Ramesh, L. Evdokimov, and R. Ensafi. Censorship in Russia, 2019. https: down-the-internet.html, 2011.
//censoredplanet.org/russia. [113] P. Winter and S. Lindskog. How the Great Firewall of China is blocking Tor. In
[85] R. Ramesh, R. Sundara Raman, M. Bernhard, V. Ongkowijaya, L. Evdokimov, 2nd USENIX Workshop on Free and Open Communications on the Internet (FOCI
A. Edmundson, S. Sprecher, M. Ikram, and R. Ensafi. Decentralized Control: 12), 2012.
A Case Study of Russia. In Proceedings of the Network and Distributed System [114] E. Wustrow, C. M. Swanson, and J. A. Halderman. Tapdance: End-to-middle
Security Symposium, NDSS 2020, San Diego, California, USA, 2020. anticensorship without flow blocking. In 23rd USENIX Security Symposium,
[86] R. Raoof, M. El-Taher, M. Tita, A. Filastò, and M. Xynou. Egypt blocks BBC 2014.
and Alhurra: Expanding media censorship amid political unrest, 2019. https: [115] E. Wustrow, S. Wolchok, I. Goldberg, and J. A. Halderman. Telex: Anticensorship
//ooni.org/post/venezuela-blocking-wikipedia-and-social-media-2019/. in the network infrastructure. In 20th USENIX Security Symposium, 2011.
[87] Reporters without Borders. Coronavirus off limits in Turkmenistan, 2020. https: [116] Xu, Xueyang and Mao, Z. Morley and Halderman, J. Alex. Internet censorship
//rsf.org/en/news/coronavirus-limits-turkmenistan. in china: Where does the filtering occur? In International Conference on Passive
[88] Reporters without Borders. Norway: Clouds in sight, 2020. https://rsf.org/en/ and Active Network Measurement, 2011.
norway. [117] M. Xynou, F. Arturo, M. Tawanda, and M. Natasha. Zimbabwe protests: Social
[89] Reporters without Borders. Turkmenistan: Ever-expanding news “black hole”, media blocking and internet blackouts, 2019. https://ooni.org/post/zimbabwe-
2020. https://rsf.org/en/turkmenistan. protests-social-media-blocking-2019/.
[90] Reuters. Polish police detain 25 after attacks on equality march, 2019. https: [118] T. K. Yadav, A. Sinha, D. Gosain, P. K. Sharma, and S. Chakravarty. Where The
//www.reuters.com/article/us-poland-lgbt-idUSKCN1UG0GH. Light Gets In: Analyzing Web Censorship Mechanisms in India. In Internet
[91] University of Oregon Route Views Project. www.routeviews.org. Measurement Conference (IMC). ACM, 2018.
[92] Q. Scheitle, O. Hohlfeld, J. Gamba, J. Jelten, T. Zimmermann, S. D. Strowes, [119] B. Zevenbergen, B. Mittelstadt, C. Véliz, C. Detweiler, C. Cath, J. Savulescu, and
and N. Vallina-Rodriguez. A long way to the top: Significance, structure, and M. Whittaker. Philosophy meets Internet engineering: Ethics in networked
stability of internet top lists. In Internet Measurement Conference (IMC). ACM, systems research. In (GTC Workshop Outcomes Paper) (September 29, 2015), 2015.
2018. [120] A. Zhang, S. Song, and J. Wang. Sequential data cleaning: a statistical approach.
[93] W. Scott, T. Anderson, T. Kohno, and A. Krishnamurthy. Satellite: Joint analysis In Proceedings of the 2016 International Conference on Management of Data, 2016.
of CDNs and network-level interference. In USENIX Annual Technical Conference [121] J. Zittrain and B. Edelman. Internet filtering in China. IEEE Internet Computing,
(ATC), 2016. 2003.
[94] P. K. Sen. Estimates of the regression coefficient based on Kendall’s tau. Journal [122] ZwNews. BREAKING: Internet shut down illegal. . . Zimbabwe High Court rules,
of the American statistical association, 1968. 2019. https://zwnews.com/breaking-internet-shut-down-illegal-zimbabwe-
[95] A. Sfakianakis, E. Athanasopoulos, and S. Ioannidis. Censmon: A web censorship high-court-rules/.
monitor. In USENIX Workshop on Free and Open Communications on the Internet
(FOCI 11), 2011.

62
Session 1A: Anonymous Routing and Censorship CCS '20, November 9–13, 2020, Virtual Event, USA

A APPENDIX: EVALUATION

Std. Deviation in censorship metric

0.7 Raw Smoothed
In this appendix, we provide an evaluation of the different anom-
0.6
aly detection techniques and the censorship smoothing. We also
provide some additional details on comparison between different 0.5

censorship measurement platforms. 0.4

0.3
A.1 Anomaly Detection Evaluation 0.2

We perform different trials with varying thresholds for various 0.1

anomaly detection techniques (§5.3.2). The MAD technique uses 0.0
the deviation from the median as an anomaly score, while the Discard Echo HTTP HTTPS DNS

likelihood model uses the likelihood of an element’s probability

in a particular distribution. The exponentially weighted moving Figure 10: Standard deviation in raw and smoothed censor-
average model calculates a weighted moving average over a sliding ship metrics–The smoothed metric is much less volatile compared
window and uses the deviation from the mean to assign anomaly to the raw censorship metric. IP censorship has similar results, but
scores. The Bitmap-based model discretizes the data into bitmaps is not shown here due to high variation in scale.
and calculates the distance between two bitmaps as anomaly scores.
Our goal is to minimize the percentage of anomalies while max-
imizing the amount of useful censorship events detected. This is events. For example, the Bitmap detection technique applied to the
difficult since there is little censorship ground truth to compare. raw censorship metric at threshold 3.1 only identifies 5 censorship
Therefore, we manually compile a list of ten key censorship events events (compared to the 7 in Table 2 when applied to the smoothed
we identified from Censored Planet data manually (described in metric).
§7.1) and observe how many of the events can be detected automat-
ically at different thresholds for each technique.
A.3 Detailed Comparison
Our evaluation is shown in Table 2. We report anomaly percent- Table 3 provides more detail on our comparison of Censored Planet
ages for a time series that is drilled down to a per-category and (03/2020) with ICLab (09/2018), OONI’s web connectivity dataset
per-country level, for which the raw number of observations is in (03/2020) and the individual remote measurement techniques (03/2020)
the order of 106 and the raw number of anomalies is in the order (§6.2). As seen from the table, not only does Censored Planet have
of 104 at the optimum level. While all of the detection techniques more coverage in terms of total number of countries, it also has
perform comparatively well, the Likelihood-based and MAD-based vantage points in all the countries in the “Not Free” category and
techniques consistently found a larger number of anomalies, prob- all but one in the “Partly Free” category of the Freedom on the Net
ably because the techniques detect smaller events in periods of 2019 report. Censored Planet also has more coverage in terms of
minimal change. The Exponentially Weighted Moving Average and raw number of ASes.
the Bitmap-based anomaly detection techniques detect compara-
tively lower number of anomalies. The Bitmap method performs B APPENDIX: RESULTS
slightly better, especially at finding most of the known censorship In this Appendix, we document results on measuring the blocking
events. Therefore, we report the top four events found using the of Tor Bridges and describe some censorship case studies other
Bitmap technique in Table 1. than the ones in §7.1. We also describe some general results.
Additionally, we observed an average overlap of 58.97 % between
comparable thresholds of the MAD, Bitmap and EWMA techniques, B.1 Blocking of Tor Bridges
indicating that a voting scheme may be used in the future to detect Upon request from Tor, we have been running custom rapid focus
the most important anomalies. The anomaly detection process is measurements testing IP reachability to Tor default bridges since
online and completely automated, although there is effort involved January 2020. The default Tor bridges are hardcoded into the Tor
in exploring causes for censorship change once the top anomalies browser and act as a valuable indicator of Tor censorship. Using
have been identified. a custom extension to Augur that allows testing connections on
different TCP ports, we tested reachability to 12 Tor bridges [105].
A.2 Censorship Smoothing Evaluation Four of these bridges were offline during the period of our measure-
In §6.3, we evaluated the high variation in raw censorship values ments. The remaining eight Tor bridges are blocked in China in
in countries with heterogeneous censorship policies, highlighting all of our measurements [38]. Tor bridges are also blocked aggres-
the importance of the smoothed representative censorship measure sively in Tanzania (seven bridges blocked), Venezuela (five bridges
we introduced in §5.2. In Figure 10, we show that the smoothed blocked) and Ukraine (five bridges blocked). Our continued testing
censorship metric is effective in reducing the volatility of the raw of reachability to Tor IPs will help discover Tor blocking patterns
censorship metric in each of our different time series. We also ob- and trends in different countries.
serve that this reduction in volatility caused by rogue vantage points
helps in obtaining a more clear signal when nationwide censorship B.2 Other Censorship Case Studies
events do occur. Applying our anomaly detection techniques on the In this section, we provide details on a few more key censorship
raw censorship metric consistently finds lower number of useful events described in Table 1.

63
Session 1A: Anonymous Routing and Censorship CCS '20, November 9–13, 2020, Virtual Event, USA

Table 2: Evaluation of Anomaly Detection techniques–The percentage of anomalies and number of events detected (out of 10).

MAD Bitmap EWMA Likelihood

Threshold % anomalies # events Threshold % anomalies # events Threshold % anomalies # events Threshold % anomalies # events
1 11.97 7 2.8 11.67 9 2.2 9.7 8 -1 17.89 8
2 9.05 6 2.9 10.38 8 2.3 8.59 6 -1.05 15.85 7
3 7.43 5 3 8.92 8 2.4 7.32 6 -1.1 14.53 5
4 6.42 4 3.1 4.79 7 2.5 3.52 5 -1.15 13.6 4
5 5.7 4 3.2 2.94 5 2.6 3 5 -1.2 12.83 4

Table 3: Comparison of scale with other censorship measure- shows the value of Cens(Smooth) (Equation 5) over time for the so-
ment platforms– Note: Censored Planet Pot. : Censored Planet cial networking category in Zimbabwe. A large increase in HTTP(S)
potential. OONI: OONI web connectivity dataset. blocking in the week of January 20 indicates the use of the SNI field
for blocking specific domains. In addition to the five social media
Platforms #AS #Country Not Partly Median Maximum domains discovered to be blocked by OONI, Censored Planet found
Free Free #ASes / #ASes / eight other domains being blocked during this period - linkedin.com,
(21) (29) country country
weibo.com, vk.com, myspace.com, foursquare.com, twimg.com,
ICLab 56 48 4 10 1 22
OONI 1,915 155 21 26 4 347 ok.ru and www.pinterest.com. These additional findings demon-
Satellite 4,713 175 21 28 5 1,067 strate the importance of testing domains on more vantage points, and
Quack 2,801 166 19 28 3 471 indicates the complementary insights Censored Planet can provide to
Hyperquack 3,872 191 19 27 7 217
Augur 314 140 17 25 2 6 existing platforms.
Censored 9,014 221 21 28 8 1,427 Although Zimbabwe’s High Court ruled on January 21st 2019
Planet
Censored 13,569 222 21 28 8 3,685
that Internet blackouts were illegal [122], we observed later in-
Planet Pot. stances of intermittent blocking of social media websites (Figure 11)
and high censorship in general. In late 2019, we observed extremely
aggressive but intermittent DNS blocking of Facebook and Insta-
gram by AS 328235 (Zimbabwe Internet Exchange). In February and
Echo HTTPS
March 2020, our Quack Echo measurements observed the blocking
Cens(smooth)

HTTP DNS
100 of 17 Social Networking websites, including Twitter, Google, and
Instagram in AS 37184 (Powertel Communications). We did not
10−1 have vantage points in AS 37184 before February 2020. Our analysis
of Zimbabwe’s continued blocking of social media domains further
10−2 illustrates the power of the longitudinal data collection and processing
Feb '19

Feb '20
Dec '18
Jan '19

Mar '19
Apr '19
May '19
Jun '19
Jul '19
Aug '19
Sep '19
Oct '19
Nov '19
Dec '19
Jan '20

Mar '20
Apr '20

of Censored Planet.

B.2.2 Blocking of News Media in Japan. In June 2019, Japan hosted

Figure 11: Social Networking Censorship in Zimbabwe– Cen- the G20 Conference for the first time [21]. The G20 conference is
sored Planet observed an increase in HTTP(S) blocking of So- a forum where 19 countries and the EU meet to discuss the global
cial Networking domains in Zimbabwe in January 2019. Censored economy and set financial regulations. Japan is noted by Freedom
Planet also detected blocking of popular Social Networking domains House to be a free country, which has resulted in many censorship
in late 2019 and 2020 using DNS and Echo measurements. studies overlooking measurements in Japan. In fact, ICLab noticed
high rates of blocking of domains in the news and media category
in Japan, but considered it as a possible false positive or localized
observation since Japan is generally thought of as a free country [7].
B.2.1 Blocking of Social Media in Zimbabwe. In January 2019, During the G20 period, we observed increased blocking of do-
protests erupted in Zimbabwe in response to skyrocketing fuel mains in the news media and E-commerce category in Japan. DNS
prices [17]. During the third week of January 2019, 12 people were blocking was observed in both categories while Echo blocking was
reportedly killed and many more protesters were wounded or ar- seen in the E-commerce category to a smaller extent. The domains
rested by the police. In response to the protests, the government being blocked during this time period included popular news do-
resorted to censorship of social media, and an entire Internet shut- mains such as online.wsj.com and washingtonpost.com under the
down in some cases [117]. As reported by OONI, five social media news media category and kickstarter.com and marketwatch.com
websites (Facebook, WhatsApp, Twitter, Instagram, and YouTube) under the E-commerce umbrella. We observed DNS blocking in
were intermittently blocked by multiple ISPs between January 14th 47 ASes (out of 51) during this week showing that the blocking is
and January 21st 2019. The report suggests blocking of HTTP con- country-wide and is not localized. The highest increase in blocking
nections to these websites. was in AS 45688 (UT-NSRG). Again, we find that Censored Planet’s
Censored Planet also detects a large increase in censorship of large scale and data processing robustness helps us uncover censorship
domains belonging to the social networking category. Figure 11 events in countries generally regarded as free.

64
Session 1A: Anonymous Routing and Censorship CCS '20, November 9–13, 2020, Virtual Event, USA

Table 4: Censorship of Different Categories. B.3 General Results

Table 4 shows the categories of domains and their overall average
Category Cens ( Smooth) Cens(Smooth) (Equation 5). Anonymization tools are at the top of
Anonymization and circumvention tools 2.19 the list, suggesting that censors are actively trying to prevent their
Foreign relations and military 1.71
Pornography 1.67
users from access content through any means necessary, and pro-
Search Engines 1.66 vides further motivation for testing reachability to circumvention
History, arts and literature 1.36 system using Censored Planet’s rapid focus capabilities. Websites
Media sharing 1.2
Social Networking 1.06 related to foreign military and pornography follow.
File-sharing 1.0 Table 5 showcases the top 5 countries and the top 3 categories
News Media 0.95 in each country having the highest Cens(Smooth) (Equation 5)
Human Rights Issues 0.72
Gambling 0.65 in each censorship method measured by Censored Planet. Our
Communication Tools 0.64 results agree with observations from other censorship measurement
Hosting and Blogging Platforms 0.63
Gaming 0.45
platforms [7, 104] but some unexpected countries (Vatican City,
Economics 0.44 Oman) enter the list because of the improved scale of Censored
Sex Education 0.44 Planet. China, Iran and Turkmenistan still dominate the list, with
Provocative Attire 0.42
E-commerce 0.39 pornography and anonymization tools being highly blocked in all
Online Dating 0.35 of these countries.
Illegal 0.33
Intergovernmental Organizations 0.29
Hacking Tools 0.28
Religion 0.24
Culture 0.24
Terrorism and Militants 0.18
LGBT 0.17
Political Criticism 0.17
Government 0.13
Hate Speech 0.11
Alcohol & Drugs 0.1
Miscellaneous content 0.1
Public Health 0.09
Environment 0.02

65
Session 1A: Anonymous Routing and Censorship CCS '20, November 9–13, 2020, Virtual Event, USA

Table 5: General results– The top five countries with the highest Cens(Smooth) (Equation 5) in each censorship method measured by
Censored Planet, and the top 3 categories blocked in each country.

Discard Echo IP
Country Category Eq 5 Country Category Eq 5 Country Category Eq 5
Turkmenistan File-sharing 40.21 Fiji Alcohol & Drugs 6.81 Cayman Islands Illegal 40.0
Media sharing 35.79 Gaming 6.69 Terrorism & Militants 37.5
Anonymization tools 31.35 Religion 6.66 Culture 22.94
Overall 11.18 Overall 6.22 Overall 11.85
China Anonymization tools 44.8 Turkmenistan Anonymization tools 12.33 Bhutan Illegal 37.5
Pornography 38.89 Social Networking 9.81 Terrorism & Militants 25.0
Terrorism & Militants 31.83 Communication tools 9.59 Culture 21.86
Overall 6.65 Overall 5.92 Overall 11.76
Oman Pornography 72.45 Oman Pornography 15.75 Guinea-Bissau Terrorism & Militants 50.0
Anonymization tools 58.73 Anonymization tools 14.55 Illegal 40.0
Terrorism & Militants 27.46 Terrorism and Militants 9.8 Hate Speech 22.51
Overall 5.08 Overall 3.96 Overall 11.25
Qatar Pornography 60.53 China Anonymization tools 6.66 Niger Terrorism & Militants 49.8
Anonymization tools 56.16 Search Engines 6.05 Illegal 33.33
Online Dating 9.35 Pornography 5.79 Culture 18.75
Overall 3.0 Overall 3.77 Overall 11.17
Iran Pornography 4.17 Qatar Pornography 9.67 Guernsey Terrorism & Militants 49.8
Provocative Attire 3.95 Anonymization tools 9.07 Illegal 49.8
History, arts & literature 3.74 Online Dating 4.38 Hate Speech 22.85
Overall 1.72 Overall 2.7 Overall 8.76
HTTP HTTPS DNS
Country Category Eq 5 Country Category Eq 5 Country Category Eq 5
Turkmenistan Anonymization tools 12.83 Vatican City Pornography 16.36 China Foreign relations & Military 49.53
File-sharing 11.26 Provocative Attire 16.3 Anonymization tools 47.72
Media sharing 10.82 Hate Speech 14.25 History, arts and literature 38.45
Overall 5.74 Overall 5.0 Overall 16.32
Comoros Gambling 10.84 Oman Pornography 12.8 Turkmenistan Anonymization tools 61.12
Pornography 10.19 Anonymization Tools 12.26 Pornography 52.35
Alcohol & Drugs 8.71 Online Dating 6.27 Media sharing 36.45
Overall 4.95 Overall 4.12 Overall 15.58
Oman Pornography 13.24 China File-sharing 7.53 Iran Pornography 45.75
Anonymization tools 12.21 News Media 6.91 Anonymization tools 44.67
Online Dating 7.4 Media sharing 6.41 Provocative Attire 29.68
Overall 4.56 Overall 3.39 Overall 14.3
Vatican City Pornography 15.12 Uzbekistan Gambling 8.44 Afghanistan Pornography 28.73
Provocative Attire 15.06 Terrorism & Militants 8.23 Anonymization tools 27.95
Hate Speech 12.99 Pornography 8.0 Provocative Attire 13.7
Overall 4.4 Overall 2.62 Overall 3.7
Uzbekistan Gambling 10.13 Turkmenistan Social Networking 8.41 Burkina Faso Provocative Attire 15.06
Terrorism & Militants 9.61 Communication tools 7.2 Online Dating 14.69
Pornography 9.41 Media sharing 6.37 Pornography 14.2
Overall 3.14 Overall 2.58 Overall 2.48

66