Deanonymisation of clients in Bitcoin P2P network

Alex Biryukov
University of Luxembourg
[email protected]
Dmitry Khovratovich
University of Luxembourg
[email protected]
Ivan Pustogarov
University of Luxembourg
[email protected]
Abstract
Bitcoin is a digital currency which relies on a distributed
set of miners to mint coins and on a peer-to-peer network
to broadcast transactions. The identities of Bitcoin users
are hidden behind pseudonyms (public keys) which are rec-
ommended to be changed frequently in order to increase
transaction unlinkability.
We present an ecient method to deanonymize Bitcoin
users, which allows to link user pseudonyms to the IP ad-
dresses where the transactions are generated. Our tech-
niques work for the most common and the most challenging
scenario when users are behind NATs or rewalls of their
ISPs. They allow to link transactions of a user behind a
NAT and to distinguish connections and transactions of dif-
ferent users behind the same NAT. We also show that a natu-
ral countermeasure of using Tor or other anonymity services
can be cut-o by abusing anti-DoS countermeasures of the
bitcoin network. Our attacks require only a few machines
and have been experimentally veried. We propose several
countermeasures to mitigate these new attacks.
1. INTRODUCTION
Digital currency based on cryptography is not a new idea [6]
but till recently it did not attract much attention. It changed
rapidly with introduction of bitcoin [11]. Bitcoin is a decen-
tralized digital currency which does not rely on a trusted is-
suing entity but rather on a peer-to-peer network with peers
minting bitcoins by brute-forcing double SHA-256 hash func-
tion. To make the money generation process computation-
ally hard, the Bitcoin protocol requires the minters to present
the hash value of a data block with new portion of Bitcoins
and new transactions to have a certain number of zeros (an
instance of the Proof-of-Work concept).
Bitcoin is now accepted as a currency by many companies
from online retailer Overstock to exotic Virgin Galactic [3].
One of its main advantages over bank transfers is its de-
centralized architecture and absence of intermediaries. This
prevents shutting it down or seizing by a government. Bit-
coin money transfers are non-refundable, reasonably fast
1
and allow to send money to any part of the world. The Bit-
coin peer network consists of homogeneous nodes and pro-
vides peer discovery and reputation mechanisms to achieve
stability. The number of Bitcoin peers is estimated to be
about 100,000 nowadays. The vast majority of these peers
1
The network sees a transaction immediately, but the re-
ceiver has to wait for 1-2 hours to be sure that there is no
double-spending.
(we call them clients), about 90%, are located behind NAT
and do not allow any incoming connections, whereas they
choose 8 outgoing connections to servers (Bitcoin peers with
public IP).
In a Bitcoin transaction, the address of money sender(s)
or receiver(s) is a hash of his public key. We call such
address a pseudonym to avoid confusion with the IP ad-
dress of the host where transactions are generated, and the
latter will be called just address throughout the text. In
the current Bitcoin protocol the entire transaction history
is publicly available so anyone can see how Bitcoins travel
from one pseudonym to another and potentially link dier-
ent pseudonyms of the same user together. A theoretical
possibility of such attack was already mentioned in the orig-
inal Bitcoin paper [11]. Since then several papers [10, 13]
showed that it is indeed possible by analysing the transaction
graph to cluster pseudonyms to dierent users. Combined
with some other sources (e.g. forum posts), the clusters
(and thus the users) can sometimes be mapped to real iden-
tities [12, 10]. Even so, these methods are not generic, and
the problem of how to tie a bitcoin address to an actual
identity remained unsolved.
Evidently, studying the entire IP trac of the Bitcoin
peers would reveal the origins of each transaction and dis-
close the identities of many users, but how much can be
achieved by an ordinary attacker with a few machines and
no access to clients behind NAT has been unclear.
Koshy et al. [9] were the rst who attempted an attack in
this direction and managed to deanonymize 1162 addresses
over the period of 5 months. Their approach, however, is
limited to the transactions that expose anomalous behaviour
like transactions relayed only once or transaction that were
relayed multiple times by the same IP. Secondly, the pro-
posed method only allows to get IP addresses of servers,
which constitute only 10% of the network, and not of the
clients. Finally, their paper does not discuss the case when
a bitcoin peer protects himself by proxying his transactions
through the Tor anonymity network.
Our contributions.
In this paper we describe a generic method to deanonymize
a signicant fraction of Bitcoin users and correlate their
pseudonyms with public IP addresses. The method explic-
itly targets the clients (i.e. peers behind NAT or rewalls)
and can dierentiate the nodes with the same public IP. Fur-
thermore, our method also handles the case when the clients
use anonymity services like Tor. If a client uses two dierent
pseudonyms during a single session, and even if they are un-
a
r
X
i
v
:
1
4
0
5
.
7
4
1
8
v
1


[
c
s
.
C
R
]


2
8

M
a
y

2
0
1
4
related in the transaction graph so that the linkage would be
totally unachievable via the transaction graph analysis [10],
our method is likely to catch it and glue the pseudonyms
together. The method is generic and might be used in other
P2P networks.
The crucial idea is that each client can be uniquely iden-
tied by a set of nodes he connects to (entry nodes). We
show that this set can be learned at the time of connection
and then used to identify the origin of a transaction.
Our attack requires only a few machines that establish a
certain number of connections by Bitcoin protocol and log
the incoming trac. In a concrete example, an attacker with
a few GB of storage and no more than 50 connections to each
Bitcoin server can disclose the senders IP address in 11%
2
of all transactions generated in the Bitcoin network. If the
attacker allows a slight DoS of the network, he may achieve
deanonymization rates up to 60%, which has been conrmed
by the experiments in the Bitcoin test network. We estimate
the cost of the attack on the full bitcoin network to be under
1500 EUR per month.
The computational power needed to disclose the sender
of a single transaction is negligible and is far smaller than
the amount of work needed to process the transaction graph
in [12, 10]. For the best of our knowledge this is the rst
attack which targets Bitcoin peers behind NAT. Our attack
does not assume any anomaly in the behaviour of peers or in
the trac and would work even if Bitcoin would encrypt the
connection. It might be applicable to other digital currencies
derived from Bitcoin.
As another interesting though unrelated to deanonymisa-
tion idea we look at how to decrease block mining diculty
by creating an alternative blockchain reality. This becomes
important since Bitcoin by design is not adaptive to rapid
drops in hash power of miners and might become necessary
in case of many miners quit mining. This is not just a hypo-
thetical case, since Bitcoin exchange rate can fall suddenly
and rapidly, making block mining unprotable.
Roadmap.
Our paper is structured as follows:
We give necessary background of how Bitcoin works
and the rules its peers follow to broadcast their ad-
dresses and transactions.
As a rst step towards deanonymization, we show how
to prohibit Bitcoin clients from using the Tor anonymity
network by exploiting Bitcoin anti-DoS protection mech-
anism (Section 3).
We show how to learn the connections of the Bitcoin
clients in Section 4.
We nally show how to identify the sender of a trans-
action (i.e. deanonymize him) in Section 5. We recover
the public IP address of the sender and further dier-
entiate clients sharing the same public IP.
We discuss how to choose parameters of the attack
and its success rate and explain our experiments on
the test network. We also propose countermeasures to
mitigate the attack.
2
The actual fraction of deanonymized transactions depends
on how stealthy an attacker wants to be.
As an extra result, we outline a strategy to lower the
diculty of the system by adding a properly selected
value to the list of checkpoints nodes hard-coded in the
client code. It can be used by the entire community
if the mining becomes unbearable and non-protable,
or by malicious administrators who want to ruin the
system (Section 8).
Ethical considerations. All vulnerabilities described in this
paper were reported to bitcoin core developers. When pos-
sible we carried out experiments in the Bitcoin test network.
To protect user privacy, we restricted from performing a full-
scale deanonymization in the real network. However, gath-
ering some statistics required us conducting experiments on
the main network, which did not cause disruption or expo-
sure of the main network.
2. BACKGROUND
In this section we provide a basic overview of the Bitcoin
system. Originating from a proof-of-concept source code
and an accompanying it white paper [11], the Bitcoin pro-
tocol never had been fully documented, and is de-facto the
functionality of the primary Bitcoin client, bitcoind [1]. In
the following text we provide only the details of the protocol
that are crucial to our research. These details are accumu-
lated from the source code of bitcoind and to a large extent
are explained in the informal wiki documentation [2].
Block chain.
Bitcoin operates on a list of blocks, the block chain. Each
block contains a header and transaction data
3
. The 80-byte
header Head contains the 256-bit hash of the previous block
Hi1, the timestamp (in seconds) Ti, the 32-bit nonce Ni
(used to generate blocks), the hash TXi of the transaction
data , and the diculty parameter di. To be valid, the
double-hash of the block header must be smaller (as an in-
teger) than a certain value, which is a linear function of the
diculty parameter:
Hi = SHA-256(SHA-256(Hi1||Ti||TXi||di||Ni||))) < f(di).
Currently it must be smaller than 2
192
, i.e. have its 64 most
signicant bits equal to zero.
The Bitcoin miners rst collect all transactions not yet
included into a block. Then they generate the header elds
and exhaustively try dierent nonces, timestamps, and other
parameters in order to obtain a valid block. They are re-
warded by 25 BTC (about $14,000 by current market rate),
which is the very rst transaction in the attached transac-
tion list. Whenever a block is created, a miner broadcasts it
to the network, so that each node attaches it into its internal
block chain.
Payers and payees of the system are identied in the blockchain
by their bitcoin addresses, or pseudonyms. A pseudonym is
the base58-encoding of the hash of the corresponding public
key. Whenever a payer wants to transfer his coins to another
user, he generates a transaction and signs it with his private
key. Signed transactions are then added to the blockchain
by miners. By checking the signature, other bitcoin partici-
pants can verify the new ownership of the coins.
3
All these conditions are strictly enforced, and a block not
conforming to them is discarded immediately.
Bitcoin P2P network
Peers of the Bitcoin network connect to each other over an
unencrypted TCP channel. There is no authentication func-
tionality in the network, so each node just keeps a list of IP
addresses associated with its connections.
To avoid denial-of-service attacks, the Bitcoin protocol
minimizes the amount of information forwarded by peers.
Valid blocks and transactions are relayed whereas invalid
blocks are discarded. Moreover, Bitcoin implements a reputation-
based protocol with each node keeping a penalty score for
every connection. Whenever a malformed message is sent to
the node, the latter increases the penalty score of the con-
nection and bans the misbehaving IP address for 24 hours
when the penalty reaches the value of 100.
Though ocial bitcoind software does not explicitly di-
vide its functionality between clients and servers, bitcoin
peers can be grouped into those which can accept incoming
connections (servers) and those which cant (clients), i.e.
peers behind NAT or rewall, etc. At the time of writing
there were about 8,000 reachable servers while the estimated
number of clients was about 100,000.
By default bitcoin peers (both clients and servers) try
to maintain 8 outgoing connections. In addition, bitcoin
servers can accept up to 117 incoming connections (thus
having up to 125 connections in total). If any of the 8 out-
going connections drop, a bitcoin peer tries to replace them
with new connections. If none of the 8 outgoing connec-
tions drop, the peer will stay connected to them until it is
restarted. In case of a client, we call the 8 nodes to which
it establishes connections entry nodes (see Fig. 1). A bit-
coin server accepts any number of connections from a single
IP address as long as the treshold for the total number of
connections is not reached.
...
...
...
...
...
Entry1
Entry2
Entry8
Client
Figure 1: Bitcoin network
Address propagation.
The bitcoin protocol implements an address propagation
mechanism to help peers to discover other peers in the P2P
network. Each Bitcoin peer maintains a list of addresses
of other peers in the network and each address is given a
timestamp which determines its freshness. Peers can re-
quest addresses from this list from each other using GETADDR
messages and unsolicitely advertise addresses known to them
using ADDR messages
4
. Whenever a bitcoin node receives an
4
One ADDR message can contain any number of address, how-
ADDR message it decides individually for each address in the
message if to forward it to its neighbours. It rst checks if
(1) the total number of addresses in the corresponding ADDR
message does not exceed 10, and (2) the attached timestamp
is no older than 10 minutes. If either of these two checks
fails, the address is not forwarded; otherwise the address is
scheduled for forwarding
5
to two of the nodes neighbours
in case the address is reachable and to one neighbour only
if it is non-reachable. An address is considered reachable
by a node if the node has a network interface associated
with same address family. Otherwise the address is marked
as unreachable. According to the current reference imple-
mentation Bitcoin nodes recognize three types of addresses:
IPv4, IPv6, and OnionCat addresses [5]. Limiting the num-
ber of neighbours to which an address is forwarded reduces
the total amount of trac in the bitcoin P2P network.
In order to choose neighbours to which to forward an ad-
dress, a bitcoin node does the following. For each of its
neighbours it computes a hash of a value composed of the
following items: address to be forwarded, a secret salt, cur-
rent day, and the memory address of the data structure de-
scribing the neighbour. The exact expression for the hashed
value is of little importance for our attacks. The only thing
which we need to emphasize is that the hash stays the same
for 24 hours. The peer then sorts the list of its neighbours
based on the computed hashes and chooses the rst entry
or two rst entries (which depends on the reachability of
the address). In the rest of the paper we call such nodes
responsible nodes for the address.
The actual transmission of the scheduled ADDR messages
does not happen immediately. Every 100 milliseconds one
neighbour is randomly selected from the list o all peers
neighbours and the queue for outgoing ADDR messages is
ushed for this node only. We call the node chosen at the
beginning of a 100 milliseconds round trickle node and the
procedure as a whole as trickling.
Consider an example on Fig. 2. Assume that node n0 gets
an ADDR message with one address A
0
from node n3 and that
node n0 schedules to forward it to nodes n1 and n2 (i.e. these
nodes are responsible nodes for address A
0
). In round 1 node
n1 is chosen as a trickle node and the address is forwarded
to this node while the delivery to n2 is still pending. After
100 milliseconds in round 2 n3 is chosen as the trickle node
thus no actual transmission happens at this stage. After
another 100 milliseconds in round 3 n2 is chosen as the trickle
node and address A
0
is nally sent to it. Choosing a trickle
node causes random delays at each hop during an address
propagation.
Finally for each connection, a bitcoin peer remembers ad-
dresses that were forwarded over this connection. Before
a peer forwards an address, it rst checks if the same ad-
dress was already sent over the connection. This history is
cleared every 24 hours. An important note is that the his-
tory of sent addresses is kept per connection and not per IP,
i.e. if a Bitcoin peer reconnects, its history will be cleared.
The total number of addresses a bitcoin peer can store is
limited by 20480. Whenever new addresses arrive at a peer
they replace old ones (according to specic rules which are
ever messages containing more than 1000 addresses are re-
jected on the remote side.
5
By scheduling a transmission we mean that the node puts
the corresponding message to the outgoing queue but does
not yet make the actual transmission.
N1
(trickle)
N2
N3
N2
N1
N3 (trickle)
Round 1 Round 2
N2
(trickle)
N1
N3
Round 3
N0 N0
N0
Figure 2: Trickling of ADDR messages
outside of the scope of this paper). In addtition when a peer
receives a GETADDR messages it sends back 23% of the number
of addresses it stores but not more than 2500 addresses.
Peer discovery.
After the startup a bitcoin peer discovers its own IP ad-
dresses, which includes not only its network interfaces ad-
dresses but also the IP address as it is seen from the Internet
(in the majority of cases for NAT users it resolves to an IP
address of the peers ISP). In order to discover the latter,
the peer issues a GET request to two hard-coded web-sites
which reply with the address. For each address obtained
by the discover procedure, the peer assigns a score. Local
interfaces initially get score 1, the external IP address gets
score of 4 (in case the external IP address coincides with one
of the local addresses the scores a summed). When a client
establishes an outgoing connection to a remote peer, they
rst exchange VERSION messages and the client advertises its
address with the highest score. The remote peer then uses
the addresses propagation algorithm described above. The
client repeats the same procedure for the remaining 7 out-
going connections.
Transaction propagation.
Forwarding a transaction from one peer to another in-
volves several steps. First the sender transmits an INVEN-
TORY message with the hash of the transactions. Second, the
receiver runs several checks on the transaction and if the
checks pass, it requests the actual transaction by sending a
GETDATA message. The sender then transmits the transaction
in a TRANSACTION message. When the receiver gets the trans-
action he advertises it to its peers in an INVENTORY message.
When a client generates a transaction he schedules
6
it for
forwarding to all of its neighbours. It then computes a hash
of a value composed of the transaction hash and a secret salt.
If the computed hash has two last bits set to zero the trans-
action is forwarded
7
immediately to all the 8 entry nodes.
6
By scheduling we mean that the node puts the transaction
into the outgoing queue of the corresponding connection.
7
More precisely the peer sends an INVENTORY message with
Otherwise a queue of a neighbour for outgoing transactions
is ushed when the neighbour becomes the trickle node (the
same as with ADDR messages). Obviously
1
4
of all transaction
are forwarded immediately in average.
When a transaction is received it is scheduled for the de-
livery to all peers neighbours as described above. As with
ADDR messages, a bitcoin peer maintains history of forwarded
transactions for each connection. If a transaction was al-
ready sent over a connection it will not be resent again. A
bitcoin peer keeps all received transaction in a memory pool.
If the peer received a transaction with the same hash as one
in the pool or in a block in the main block chain, the received
transaction is rejected.
3. DISCONNNECTING FROM TOR
In this section we explain the rst phase of our attack.
We show how to prohibit the Bitcoin servers to accept con-
nections via Tor and other anonymity services. This results
in clients using their actual IP addresses when connecting
to other peers and thus being exposed to the main phase of
our attack, which correlates pseudonyms with IP addresses.
This phase is quite noticeable, so a stealthy attacker may
want to skip it and deanonymize only non-Tor users.
In the further text we discuss Tor, but the same method
applies to other anonymity services with minor modica-
tions. Briey, the Tor network [7] is a set of relays (5397 for
the time of writing) with the list of all Tor relays publicly
available on-line. Whenever a user wants to establish a con-
nection to a service through Tor, he chooses a chain of three
Tor relays. The nal node in the chain is called Tor Exit
node and the service sees the connection as it was originated
from this Tor Exit node.
To separate Tor from Bitcoin, we exploit the bitcoin built-
in DoS protection. Whenever a peer receives a malformed
message, it increases the penalty score of the IP address
from which the message came (if a client uses Tor, than the
message will obviously come from on of the Tor exit nodes).
When this score exceeds 100, the senders IP is banned for 24
hours. According to the bitcoind implementation, there are
many ways to generate a message which would cause penalty
of 100 and an immediate ban, e.g. one can send a block with
empty transactions list (the size of such a message is 81
bytes). It means that if a client proxied its connection over
a Tor relay and sent a malformed message, the IP address
of this relay will be banned.
This allows to separate any target server from the entire
Tor network. For that we connect to the target through as
many Tor nodes as possible. For the time of writing there
were 1008 Tor exit nodes. Thus the attack requires estab-
lishing 1008 connections and sending a few MBytes in data.
This can be repeated for all Bitcoin servers, thus prohibit-
ing all Tor connections for 24 hours at the cost of a million
connections and less than 1 GByte of trac. In case an IP
address of a specic Bitcoin node can be spoofed, it can be
banned as well.
As a proof of concept we used the described method to
isolate our bitcoin node from a set of Tor exit relays.
Possible countermeasures.
It is desirable to allow the Bitcoin peers to use Tor and
still to keep some blacklisting capability. We suggest making
the hash of the transaction.
every connection time- or computation-consuming to radi-
cally increase the attack cost. For instance, any peer that
initiates a connection might be required to present some
proof-of-work, e.g., a hash of its IP, the timestamp, and the
nonce that has a certain number of trailing zeros. If we re-
quire 32 zero bits, then to separate a single peer from the
Tor network would cost about 2
45
hash computations, which
takes several days on a modern PC.
One may argue that some Bitcoin pools are powerful enough
to aord that many hash calls. However, the vast majority of
pools computing power is contained in custom-built ASIC
miners, which implement only a specic instance of SHA-
256 and can not be recongured for another hash function,
say, SHA-3. The exact fraction of GPU and CPU computing
power is unknown, but at the time when these architectures
were dominant, the total computing power was by several
orders of magnitude smaller than now.
4. LEARNING TOPOLOGY
Suppose that we have ruled out the case that the Bit-
coin users, which we deanonymize, use Tor. Now we target
clients, i.e. nodes that do not accept incoming connections,
but have 8 outgoing connections to entry nodes. In this sec-
tion we show how to learn these entry nodes.
The method is based on the fact that whenever a client
C establishes a connection to one of its entry nodes, it ad-
vertises its address Ca as it is seen from the Internet (see
section 2). If the attacker is already connected to an entry
node, with some probability (which depends on the num-
ber of the attackers connections) the address Ca will be
forwarded to him. This suggests the following strategy:
1. Connect to W Bitcoin servers, where W is close to the
total number of servers.
2. For each advertised Ca, log the set E

of servers that
forwarded Ca to attackers machines and designate it
as the entry node subset E

Ca
.
There are two problems with this method. First, the entry
node might send the clients address to some non-attackers
peer. Second a client does not connect to all his entry nodes
simultaneously, but there is a time gap between connections.
In both cases, the advertised address reaches attackers ma-
chines via peers that are not entry nodes, which yields false
(noisy) entries in E

Ca
.
Noise reduction technique.
Our strategy of ltering noise assumes that either the
clients IP was already used in the Bitcoin network, which
is quite common for the clients behind NAT or the clients
public IP is contained in a known list of IP addresses (e.g.
within an IP range of a major ISP) which an attacker can
use. If an attacker knows Ca, he restricts its propagation
using the following fact:
If the address had already been sent from A to B, it
will not be forwarded over this connection again;
This suggests broadcasting Ca (or all the addresses under
investigation) to all servers we are connected to. We sug-
gest repeating this procedure every 10 minutes (see details
below), though there could be other options. The adversary
expects that when the client reconnects, the entry nodes
will forward Ca to him, and even if they dont, the address
propagation will stop before it reaches the adversary via a
non-entry node.
Eventually the attacker obtains the fraction p
addr
of clients
entry nodes. The exact value of p
addr
depends on the num-
ber of attackers connections, and it is computed for some
parameters in Section 7.1. For instance, if an attacker es-
tablishes 35 connections to each potential entry node, which
all had 90 connections beforehand, then he identies 4 entry
nodes out of 8 on average.
Here are some details. When the attacker advertises the
Ca, each bitcoin server chooses two responsible nodes to for-
ward the address. The attacker than establishes a number
of connections to each server in the network hoping that her
nodes will replace some of the responsible nodes for address
Ca. When client C connects to one of its entry nodes e1,
it advertises its address. If one of attackers nodes replaced
one of the responsible nodes, then the attacker will learn
that client C might be connected to node e1. If the respon-
sible nodes did not change address Ca will not be propagated
further in the network.
Since the attacker advertised Ca to node e1, responsible
nodes of e1 might be replaced by some non-attacker nodes
and the attack might fail. In Section 7 we show that the
probability of this event is actually quite low given that the
attacker re-sends its list of addresses frequently enough.
5. DEANONYMIZATION
We have prohibited Bitcoin servers from accepting Tor
connections and showed how to nd the entry nodes of clients.
Now we describe the main phase of the deanonomyzation at-
tack.
The main phase consists of four steps:
1. Getting the list S of servers. This list is regularly
refreshed.
2. Composing a list C of Bitcoin clients for deanonymiza-
tion.
3. Learning entry nodes of clients from C when they con-
nect to the network.
4. Listening to servers from S and mapping transactions
to entry nodes and then to clients.
Eventually we create a list I = {(IP, Id, PK)}, where IP
is the IP address of a peer or its ISP, Id distinguishes clients
sharing the same IP, and PK is the pseudonym used in a
transaction (hash of a public key). Let us explain the steps
in detail.
Step 1. Getting the list of servers.
This phase of the attack is rather straightforward. An
attacker rst collects the entire list of peers by quering all
known peers with a GETADDR message. Each address P in
the response ADDR message can be checked if it is online by
establishing a TCP connection and sending a VERSION mes-
sage. If it is, P is designated as a server. An attacker can
initiate the procedure by querying a small set of seed nodes
and continue by querying the newly received IP addresses.
The adversary establishes m connections to each server (we
suggest 50 for the size of the current Bitcoin network).
Step 2. Composing the deanonymization list.
The attacker selects a set C of nodes whose identities
he wants to reveal. The addresses may come from various
sources. The attacker might take IP ranges of major Inter-
net service providers, or collect addresses already advertised
in the bitcoin network. Finally, she might take some entries
from the list of peers she obtained at Step 1.
Step 3. Mapping clients to their entry nodes.
Now the attacker identies the entry nodes of the clients
that are connecting to the network. Equipped with the list
C of addresses, the attacker runs the procedure described in
Section 4. Let us estimate how many entry nodes are needed
to uniquely identify the client.
Let us denote the set of entry nodes for P by EP . We
stress that it is likely that EP
1
= EP
2
even if P1 and P2 share
the same IP address. For each P advertising its address in
the network the attacker obtains a set of E

P
EP . Since
there are about 8 10
3
possible entry nodes out of 10
5
total
peers (servers and clients together), the collisions in E

P
are
unlikely if every tuple has at least 3 entry nodes:
10
5
10
5
(8 10
3
)
3
1.
Therefore, 3 entry nodes uniquely identify a user, though
two nodes also do this for a large percent of users.
An attacker adds EP to its database and proceeds to Step
4.
Step 4. Mapping transactions to entry nodes.
This step runs in parallel to steps 1-3. Now an attacker
tries to correlate the transactions appearing in the network
with sets E

P
obtained in step 2. The attacker listens for
INVENTORY messages with transaction hashes received over all
the connections that she established and for each transaction
T she collects RT the rst q addresses of bitcoin servers
that forwarded the INVENTORY message. She then compares
E

P
with RT (see details below), and the matching entries
suggest pairs (P, T). In our experiments we take q = 10.
There could be many variants for the matching procedure,
and we suggest the following version.
The attacker composes all possible 3-tuples from all
sets E

P
and looks for their appearances in RT . If there
is a match, he gets a pair (R, T);
If there is no match, the attacker consider 2-tuples and
then 1-tuples. Several pairs {(Pi, T)} can be suggested
at this stage, but we can lter them with later trans-
actions.
We made a bunch of experiments and collected some statis-
tics to estimate the success of the attack. Even the rst step
is quite powerful. In our experiments on the testnet we es-
tablished 50 connections to each server, obtained 6 out of 8
entry nodes on average, and the 3-tuples were detected and
linked to the client in 60% of transactions (Section 6). In
the real network, where we can establish fewer connections
on average, our pessimistic estimate is 11% (Section 7), i.e.
we identify 11% of transactions.
The 2-tuples may suggest several pairs. Each client has
2
5
2-tuples of entry nodes, whereas the top-10 suggests 2
5.5
2-tuples. The matching probability is 2
26
, which implies
that the top-10 suggests 2
16.5+10.526
= 2 clients on average
on the 2-tuple rule. The probability for the right client to
be detected we estimate as 0.28 in Section 7, which means
that each transaction suggests two clients, but only in 28%
cases the right one is among those two.
Remark 1.
Step 4 of the attack depends on that some entry nodes of
a client are among the rst to forward the INVENTORY message
with the transactions hash. The intuition behind it is that it
takes a number of steps for a transaction to propagate to the
next hop. Fig. 3 shows steps that are required for a trans-
action to be propagated over two hops and received at peer
A. When a transaction is received by a node it rst runs a
number of checks and then schedules the transmission. The
actual transmission will happen either immediately (for 25%
of transations) or with a random delay due to trickling (see
Section 2). The time needed for an INVENTORY message to
be forwarded to the attackers node through node Entry is
the sum of propagation delays of 4 messages (2xINVENTORY,
1xGETDATA, 1xTRANSACTION) plus the time node Entry needs to
run 16 checks and possibly a random trickling delay. On the
other hand the time needed for the same INVENTORY message
to be forwarded to the attackers node through peer A con-
sists of 7 messages (3xINVENTORY, 2xGETDATA, 2xTRANSACTION),
32 checks, and two random delays due to trickling. Finally
since the majority of connections to a peer are coming from
clients, one more hop should be passed before the trans-
action reaches an attackers node through a wrong server.
Measurements of transaction propagation delays are given
in Appendix C.
16 Checks
[Trickling]
Inventory(tx_hash)
getdata(tx_hash)
Transaction
Client Entry Peer A
Inventory(tx_hash)
getdata(tx_hash)
Transaction
Figure 3: Steps necessary to forward a transaction
Based on this we expect that if a transaction generated by
a client is forwarded to the entry nodes immediately, the en-
try nodes will be the rst nodes to forward the transaction.
In case when the transcation was sent sequentially with 100
ms between transmissions we still expect a fraction of en-
try nodes to be among the rst 10 to forward corresponding
INVENTORY message to one of the attackers nodes. This frac-
tion obviously depends on the propagation delay between
bitcoin peers. The higher the propagation delay the less sig-
nicant becomes delay of 100 ms in trickling. For example
if the propagation delay is 300 ms between the client and
each entry node its likely that 3 entry nodes will be among
the rst to forward the INVENTORY message (given that the
attacker has enough connections to bitcoin servers).
Remark 2.
The attack presented in this section requires from an at-
tacker only to be able to keep a signicant number of connec-
tions to bitcoin servers without sending large amount data.
In order to make the attack less detectable an attacker might
decide to establish connection to a given bitcoin server from
dierent IP addresses, so that all connection look like they
came from dierent unrelated clients. The same set of IP
addresses can be used for dierent servers.
Remark 3.
The technique considered in the section provides unique
identication of bitcoin clients for the duration of a session,
and thus if a client makes multiple transactions during one
session they can be linked together with very high proba-
bility. Note that this is done even if the client uses totally
unrelated public keys/bitcoin wallets, which have no rela-
tion in the bitcoin transaction graph and thus such linkage
would be totally unachievable via transaction graph analy-
sis [10, 13]. Moreover we can easily distinguish all the dif-
ferent clients even if they come from the same ISPs, hidden
behind the same NAT or rewall address.
Countermeasures.
As a possible countermeasure against client de-anonymization
we propose to change the client octet every transaction and
add some random delay after the transaction (to avoid tim-
ing linkability attack). This will remove likability of transac-
tions and will also prohibit distinguishing of dierent clients
from the same ISP. This however will not prevent the at-
tacker from learning the ISP of the client.
6. EXPERIMENTAL RESULTS
As a proof of concept we implemented and tested our at-
tack on the Bitcoin testnet. We did not perform a deanonymi-
sation attack on real clients for ethical reasons. For our ex-
periments we built our own Bitcion client, which included
functionality specic for our attack sending specic bit-
coin messages on request or establishing various numbers of
parallel connections to the same bitcoin server, etc. When
imitating clients we used the main Bitcoin client. In order
to periodically get the list of all running bitcoin servers we
used an open source crawler [4].
For the time of experiments (May 2014) the number of
running bitcoin servers in the testnet uctuated between 230
and 250, while the estimated average degree of the nodes was
approximately 30. In our experiments we were imitating
several dierent users connecting to the testnet from the
same ISPs IP address and from dierent ISPs at dierent
times. As an attacker we added 50 additional connections to
each bitcoin server. For each experiment in the rst phase
of the attack we propagated clients addresses in the testnet
10 minutes before they started to send their transactions.
In total we (as clients) sent 424 transactions.
In the rst experiment we conrm our expectations that
transactions are rst forwarded by entry nodes and analyse
the number of entry nodes that were among the rst 10 to
forward the transactions (i.e. we assume that the attacker
correctly identied all entry nodes). We splitted all trans-
actions into two sets: the rst set contains 104 transactions,
which were forwarded to the entry nodes immediately; the
second set contains all other 320 transactions (i.e. for which
trickling was used). Fig. 4 shows the number of entry nodes
that were among the rst 10 to forward the transaction to
the attackers nodes for these two sets. As expected if a
transaction was immediately forwarded to all entry nodes
the attacker was able to catch three or more of them in
99% of cases. In case of transactions from the second set,
the attacker was able to catch 3 or more entry nodes in
70% of cases. We also observed that for the majority of
transactions the rst two nodes to forward the transaction
to the attacker were the entry nodes.
0
5
10
15
20
25
30
0 1 2 3 4 5 6 7 8
%
,

t
r
a
n
s
a
c
t
i
o
n
s
Intersection of 10 first senders and entry nodes
Transmitted immediately
Trickled out
Figure 4: Intersection of top-10 senders and entry
nodes
In the second experiment we executed all steps of the at-
tack. In our experiment each client was successfully uniquely
identied by his own set of entry nodes and on average we
identied 6 entry nodes for each client. Assuming that 3
entry nodes is enough for unique identication of a client
we correctly linked 59.9% of all transactions to the corre-
sponding IP address by matching entry nodes of clients and
rst 10 bitcoin servers which forwarded the transaction. We
correctly glued together all transactions of the same client
which were made during one session.
In a bit more conservative setting we added only 20 ad-
ditional nodes in which case we successfully deanonymised
41% of our transactions.
7. ANALYSIS
The success rate of the attack presented above depends
on a number of parameters, among which the most impor-
tant is the fraction of attackers connections among all the
connections of clients entry nodes. The fewer the number of
connections of entry nodes are, the more connections the at-
tacker can establish and the higher chance is to deanonymise
the client. In this section we analyze each step of the attack
and compute success rates for some parameter sets.
7.1 Number of connections to servers
Both mapping client to entry nodes and mapping entry
nodes to transactions depends on the number of connections
the attacker can establish to the bitcoin servers. Assuming
the entry node had n connections and the attacker added
m new connections, thus the total number of connections is
N = n+m, the probability to receive the address at the rst
hop is p
addr
(n, N) = 1
n
N

n1
N1
. For a transaction which
was not forwarded immediately to the peers neighbours the
probability that one of attackers nodes is chosen as trickle
node in the rst round is ptx =
m
N
. For n = 50, m = 50,
p
addr
= 0.75 and ptx = 0.50. For n = 90, m = 35, p
addr
=
0.49 and ptx = 0.28. The number of connections that the
adversary can establish is limited by the total number of 125
connections a bitcoin peer can have by default.
In order to see how many open connection slots bitcoin
peers have we conducted the following experiment in April
2014. For each bitcoin server that we found we tried to es-
tablish 50 parallel connections and check the actual number
of established connections
8
. Fig. 5 shows the distribution of
number of established connections. The experiment shows
60
65
70
75
80
85
90
95
100
0 10 20 30 40 50
%

o
f

b
i
t
c
o
i
n

s
e
r
v
e
r
s
Number of open connection slots
Figure 5: Distribution of open slots
that 60% of peers allow 50 connections or more, and 80% of
bitcoin peers allowed up to 40 connections. Note that even
if sucient number of connection cannot be established to a
bitcoin peer immediately they can be established in longer
term since many bitcoin clients will eventually disconnect
and thus allow new connections (according to an example
disconnection rate as shown in Fig. 7 it might take several
hours, but once an attacker got the required number of con-
nections she can keep them as long as needed). Also note
that bitcoin servers allow any number of connections from a
single IP address.
Finally the attacker does not send much trac over the
established connections but rather listens for messages. In-
coming trac is normally free of charge if one rents a server.
Thus in spite of the large number of connections that an at-
tacker needs to establish the attack remains very cheap.
7.2 Estimating false positives
Assume that some of the steps of that attack fail. Then
the rst 10 peers to report the transaction to the attacker
will be some random bitcoin peers. If there is no 3-subset
of these 10 that match some entry node set, then such a
transaction is marked as unrecognized by an attacker. The
probability that nodes accidentally match any set of bitcoin
entry nodes (we call this a collision) for a given transaction
is
pc =

10
3

10
3

1
N

3
where N is the number of bitcoin servers in the network.
Given that there are about 8000 bitcoin servers and 100,000
8
We did not try establish more than 50 connections in order
not to degrade the Bitcoin network performance.
bitcoin clients, the number of incorrectly assigned transac-
tions is negligible.
We now estimate the probability that an attacker adds a
wrong entry node to the set of entry nodes of a particular
client (we recall that according to the address propagation
mechanism after receiving an address a peer forwards it to
only two randomly chosen responsible nodes). For this to
happen, one or more entry nodes should forward the clients
address Ca over one of non-attackers connections, whence
(since the attacker periodically propagates the clients ad-
dress) at least one of responsible nodes for address Ca should
change on an entry node after the attacker last propagated
Ca.
In order to estimate this probability we collected statistics
from our bitcoin peer for 60 days from March 10 till May
10 2014. We collected information about 61,395 connections
in total. Assume that the attacker propagated Ca at time
t0, the probability that a responsible node will be dierent
at time t1 = t0 + t depends on the number of new con-
nections the entry node has at t1 and number of nodes that
disconnected since t0. Fig. 6 shows probability density func-
tion of the number of new connections (i.e. the incoming
connections rate) for dierent values of t.
0
10
20
30
40
50
60
70
80
90
0 1 2 3 4 5 6 7 8
%

o
f

e
x
p
e
r
i
m
e
n
t
s
Number of new connections
t=60 seconds
t=120 seconds
t=180 seconds
t=600 seconds
t=1200 seconds
Figure 6: Probability density of number new con-
nections
Fig. 7 shows probability density function of the number of
disconnection (i.e. connection close rate) for dierent values
of t.
We use these distributions to simulate the address propa-
gation and calculate the probability that the clients address
is forwarded by an entry node over a non-attackers link af-
ter time t after the attacker sent this address over the
network. We obtained probabilities for dierent number of
attackers and non-attackerss connections and for each con-
nection setting and each t we executed 10,000 runs of the
model. Fig. 8 shows the obtained probabilities. The number
of attackers connections is denoted by m and the number
of non-attackers connections by n.
As expected, the more connection a node has the less prob-
able that the responsible nodes for an address will change
after t. Another observation is that the probability of a
node to forward the clients address over one of the non-
attackers connections depends on the total number of con-
nections rather than on the fraction of attackers connec-
tions. From Fig. 8 we conclude that resending client ad-
0
10
20
30
40
50
60
70
80
90
0 1 2 3 4 5 6 7 8
%

o
f

e
x
p
e
r
i
m
e
n
t
s
Number of lost connections
t=60 seconds
t=120 seconds
t=180 seconds
t=600 seconds
t=1200 seconds
Figure 7: Probability density of number lost con-
nections
0
1
2
3
4
5
6
7
8
0 200 400 600 800 1000 1200
%

o
f

w
r
o
n
g

p
r
o
p
a
g
a
t
i
o
n
s
Delay, seconds
n=50, m=20
n=50, m=50
n=60, m=20
n=90, m=35
Figure 8: Percentage addresses forwarded by entry
node over non-attacker connections
dresses over the bitcoin network every 10 minutes seems to
be a reasonable choice. Also note that even if a clients ad-
dress was forwarded over a non-attackers link, the further
propagation of the address will likely stop at the next hop.
7.3 Overall success rate
The success rate Pc of the attack depends on a number of
characteristics of the real network.We propose the following
method to estimate it. First, we assume that the attacker
establishes all possible connections to Bitcoin servers. From
the data used in Figure 5, we estimate the average value
p
Avg
addr
of the parameter p
addr
. We did not establish more
than 50 connections to avoid overloading servers, and we
take a pessimistic estimation that 50 is the maximal number
of attackers connections. This yields
p
Avg
addr
0.34.
Then we assume that both the testnet and the mainnet ex-
hibit similar local topology so the probabilities P3(L) for
the number L of entry nodes being in top-10 are almost the
same (Figure 4). We calculate the probabilities P1(R) for
the number R of entry nodes being detected out of 8 as
a function of p
Avg
addr
. Then we compute the total probability
that the adversary detects at least M3 = nodes among those
appeared in top-10, and we get the following estimation (de-
tails in Section A):
Psuccess(3) 0.11.
When we restrict to 2-tuples, the success rate increases to
0.35.
In the testnet we managed to achieve p
Avg
addr
= 0.86 and the
success rate for M = 3 being close to 60%. An attacker may
achieve such high rates if he rst saturates servers connec-
tions and then gradually replaces the expired connections
from other nodes with his own ones. However, this may
cause degradation of QoS as some clients will be unable to
connect to all their entry nodes.
Thus a careful attacker that followes the 3-tuple rule only
and establishes 50 connections at maximum to each server
can catch about 11% of transactions generated by clients.
Given 70,000 transactions per day, this results in 7,700 trans-
actions per day. This also means that a user needs to send
9 transactions in average in order to reveal his public IP
address.
8. ALTERNATIVE REALITY
In this section we show how to create and maintain an
alternative block chain while keeping the Bitcoin protocol,
existing wallets and transactions untouched. This procedure
can be used by the Bitcoin community if the current di-
culty becomes too high and impossible to sustain. It can
also be an attack vector for a malicious admininstrator of
the Bitcoin code project.
For motivation, suppose that the mining stops being prof-
itable so that the most powerful miners quit the network in
order to stop paying large electricity costs. As a result, the
total computational power drops and the network outputs
blocks at a slower rate.
The Bitcoin protocol is quite reactive to the rise of the
hashrate, but has large latency in the opposite case. The
maximum diculty drop is the factor of 4 and requires the
2016 blocks to be produced within at least 8 weeks. Not
only it is a long recovery period, but also the network might
be so slow that the necessary number of blocks might not be
produced at all. Our solution creates an alternative reality
with lower diculty and the same transaction history within
a small timeframe.
Block construction rules.
It may happen that distinct miners create blocks almost
simultaneously, which is called a fork. In this case the net-
work temporarily splits into two parts, which try to generate
a valid block at their own end of the fork. When a new block
is found by either part, a higher diculty principle applies:
the chain with higher total diculty
9
is accepted, and the
one with lower is discarded. As a result, a new block at ei-
ther end of the fork yields a chain with higher diculty, and
the chain is accepted by all the peers. Due to this short term
uncertainty it is recommended to wait at least 6 blocks (i.e.
about 60 minutes) as a conrmation that the transaction be-
came part of the block chain. To further x the block chain,
the administrators of bitcoind routinely hard-code hashes of
9
The real formula is complicated, but the total computa-
tional complexity of generating the chain is a good approx-
imation.
some blocks into the client (currently 12 checkpoint blocks,
on average every 25,000-th, are hard-coded) code.
The Bitcoin protocol aims to sustain a constant rate of
block production. Every 2016 blocks (about two weeks) the
diculty is recalculated so that blocks are generated every
10 minutes. The exact rules are as follows:
For block Xi, i = 2016k, the diculty is equal to that
of Xi1;
For block Xi, i = 2016k, we extract the time T2 of
block Xi1 and the time T1 of block Xi2016. Let the
time dierence T = T2 T1 be expressed in days.
Then the diculty is multiplied by
14
T
. The protocol
also enforces that the multiplier can not exceed 4 or
be smaller than 0.25.
Bitcoin enforces a number of additional restrictions to dis-
courage malicious miners to play with timestamps for their
own benet. The following two rules are important for us:
The timestamp of Xi can not be older than the me-
dian (i.e., the middle element of the sorted array) of
11 previous timestamps.
The diculty di of Xi can not be lower than the hy-
pothetical diculty yielded by reducing the last check-
point diculty by the factor of 4 every 8 weeks, i.e. the
minimal diculty that is possible if the network slows
down.
Alternative block chain.
Alternative chain is constructed as follows. First, we select
the rst block Xi after the last checkpoint such that 2016
divides i: i = 2016k, i.e. the diculty is recomputed at
this point. We create an alternative block with the same
transactions but the date changed to the current date, which
will decrease the diculty of the subsequent blocks by the
factor of 4. The next 2015 blocks we create with arbitrary
times, possibly immediately one after another, with Xi+1
and later possibly close to Xi1.
The date of block Xi+2016 we set again to the current time
so that the total diculty would drop as much as possible.
The next blocks will be again older than Xi+2016. We re-
peat this procedure further and further until the resulting
diculty contradicts the diculty of the checkpoint. If Tc is
the date (in days) and Qc is the diculty of the last check-
point the client has in memory, T is the date and Q is the
diculty of the processed block, than the lower bound is
Q
Qc
2
TTc
28
.
Currently, a new checkpoint is added every 25000 blocks,
which amounts to the period of about 140 days with the
current diculty increase rate. Therefore, the diculty may
drop by the factor of 2
10
compared to the previous check-
point.
As a result, we create an alternative reality where all the
participants have the same balance. However, the new chain
is not accepted by clients since it would have the smaller to-
tal diculty compared to the original chain. To nish the
switch to the new reality, a new checkpoint must be chosen
on the new chain and distributed among the clients. Alter-
natively, high-diculty blocks can be added to the beginning
Figure 9: Schematic view of the alternate reality
creation over 8064 blocks. Vertical axis denotes the
block index, horizontal axis denotes the timestamp
(in days) and the diculty.
of the alternate chain to make it more dicult than the orig-
inal one. Higher granularity achieved by lower diculty at
the end of the alternative chain would allow to surpass the
original chain even if the last checkpoint is not set.
Let us estimate the amount of computational power needed
for this operation. Suppose that we have waited for 25000
blocks after the last checkpoint. This occurred in Dec 15th,
2013 with the block 275000, with the checkpoint block 250000
generated on August 3d, i.e., 134 days before. It has di-
culty smaller by the factor of 30, let us denote it by D. In
turn, the diculty in our new history can be even lower by
approximately 2
134
28
30. To obtain that, we would have to
create 2016 blocks with diculty D/4 and 2016 blocks with
diculty D/16. The other 23000 blocks must be created
with diculty D/30. This amounts to about 1400 blocks
with diculty D, or less than 50 blocks with current dif-
culty. This means that a mining pool with only 10% of
the network computational power would need only 3 days
to make this happen.
9. FURTHERLEARNINGTHETOPOLOGY
In this section we continue learning the topology of the
Bitcoin network and show how to learn server-to-server con-
nections. Bitcoin peers share information only about other
peers they know, but not about their direct connections. In
this section we rst provide a method to estimate a nodes
degree (the number of connections) and then show how to
determine to which servers it is connected to.
Some of the attacks exploit the following feature of ad-
dress propagation mechanism. Each forwarded address is
accompanied with a timestamp. If this timestamp is older
than 10 minutes, the address is not retransmitted anymore.
Thus in order to avoid false positives in some attacks de-
scribed below an attacker should set the timestamps of the
marker addresses to values close to 10 minutes to the past.
9.1 Estimating number of connections
Our rst method is based on the way a bitcoin peer for-
wards addresses received in ADDR messages (see section 2 for
more details). Suppose that a bitcoin node A is connected
to k nodes P1, P2, . . . , Pk. We show now how to estimate k.
First, we connect our node X to peer A and send it a set
of fake
10
marker addresses S = {ip1, ip2, . . . , ipn} in por-
tions of 10 addresses per ADDR message. At the same time
we listen (either on the same connection or on a separate
connection) for received marker addresses. As instructed by
the Bitcoin protocol, node A forwards marker addresses to
its peers including X. As the number of marker addresses
increases the number of addresses received by X converges
to
2
1+k
if marker addresses are considred reachable by A or
1
1+k
otherwise, from which we estimate k. We increase the
accuracy by 1) running several listening nodes, 2) repeating
the same experiment several times. We can reuse the same
marker addresses: when we reconnect our listening nodes,
peer A revokes their histories and allocates new data struc-
tures for this nodes. We note that all connections can be
established from the same IP address.
We implemented this method and carried out several ex-
periments. Our target nodes have 10, 30, 70, or 100 connec-
tions. For dierent number of connections we used dierent
number of marker addresses and dierent number of listen-
ing connections (2, 3, 7, and 10 correspondingly). For each
number of connections we conducted a series of experiments;
Table 1 shows ve random runs for each series.
9.2 Determining connections between servers
The method to determine connections of a bitcoin servers
is similar to the method of estimating the nodes degree. It
is based on sending marker addresses to a peer which should
then forward them to its neighbours. The number of bitcoin
servers is estimated to be about 10% of the total number of
Bitcoin peers. Note however that clients can connect to the
bitcoin network only through connecting to servers which
means that all bitcoin messages even those generated by
clients should travel along at least one link between two
bitcoin servers. We rst describe a probabilistic method to
determine if two given peers A and B are connected which
consists of two phases.
During the rst phase the attacker estimates the number
of connections of peer A. This number is used to compute the
number of marker addresses that will be forwarded to the
peers neighbours. In the second phase the attacker chooses
a set of fake marker addresses S = {ip1, ip2, . . . , ipn} and
sends them in ADDR messages to peer A (10 addresses per
message). She then sends GETADDR messages to peer B. If the
number of marker addresses known to B corresponds to the
number estimated in the rst phase, node B is marked as a
neighbour.
An attacker can enhance this method to reveal a peers
connections by applying it to each node in the list of run-
ning bitcoin nodes (this information is available to the at-
tacker since running nodes advertise their addresses). This
is easily parallelized since the attacker needs to send marker
addresses just once. The drawbacks of the method is that it
does not allow to reveal connections to nodes which dont ac-
cept incoming connection (e.g. located behind a NAT) since
an attacker cannot send GETADDR messages to such nodes.
Bitcoin network discovery protocol is designed in such a
way so that newly advertised addresses should be delivered
10
By fake me mean that no bitcoin nodes are running at these
addresses.
to the majority of the nodes. Thus one of the key ingredients
of the method proposed in this section is how to reduce the
propagation radius. This is achieved by that each forwarded
address is accompanied with a timestamp. If this timestamp
is older than 10 minutes, the address is not retransmitted
anymore. Thus in order to avoid false positives an attacker
should set the timestamps of the marker addresses to values
close to 10 minutes to the past.
We implemented the method and carried out several ex-
periments with our own bitcoin nodes which had 59, 53, 73,
and 81 connections. As a list of candidates we used 400 ran-
domly chosen running bitcoin nodes plus the nodes current
connections. The results of the experiments are summarized
in Table 9.2. In all our experiments we had zero false posi-
tives.
In order to estimate how probable false positives are we
chose 30 random bitcoin nodes and sent them marker ad-
dresses with timestamps set to 9 minutes 58 seconds in the
past relative the nodes adjusted time. For each node we
generated a unique set of 1000 marker addresses. At the
same time we were monitoring for received addresses at two
of our nodes which had 83 and 85 connections. At the end
of the experiment no marker addresses arrived at our nodes
which indicated that false positives are quite unlikely.
Connections Not behind
NAT
Candidates Discovered
59 25 459 25
53 22 453 22
73 8 473 8
81 17 481 17
Table 2: Discovering bitcoin node connections
Finally in order to estimate the number of GETADDR mes-
sages sucient to learn addresses known to a peer we adopt
a nite state discrete time Markov Chain model [8]. Each
state in the model represents number of addresses learned
by the attacker. At each step the attacker sends a GETADDR
message and gets back 2500 random addresses from the to-
tal of maximum 20480 (note that some of those addresses
may already be known to the attacker from the previous
requests). The chain has one absorbing state which is all
addresses are known to the attacker. By computing funda-
mental matrix we get the average number of transitions be-
fore the absorbing state is reached which corresponds to the
number of messages the attacker needs to send. If the max-
imum number of addresses stored at node is 20480, it takes
in average approximately 80 GETADDR messages to learn all
those addresses. Indeed, the probability for a single address
to not be discovered is upper bounded by

2500
20480

80

1
30000
.
This estimation shows however an upper bound of the num-
ber of GETADDR messages. Our experiments showed that it
is sucient to retrieve 5 ADDR message from a peer in or-
der to conrm that a connections exists, which signicantly
reduces the number of GETADDR messages.
10. CONCLUSION
We have presented the rst method that correlates pseudonyms
of Bitcoin users behind NAT with the public IP address of
Real connections (k) Markers sent
Estimates using our formula
Try #1 Try #2 Try #3 Try #4 Try #5 Average
10 500 10.69 9.57 9.34 10.6 11.38 10.32
30 1000 31.92 30.88 35.47 36.26 30.48 33
70 1000 72.92 76.84 70.65 64.16 77.36 72.39
100 2000 102.63 109.12 104.27 103.28 95.66 103
Table 1: Estimation of number of connections
the host where the transaction is generated. The crucial
idea of our attack is to identify each client by an octet of
outgoing connections it establishes. This octet of bitcoin
peers (entry nodes) serves as a unique identier of a client
for the whole duration of a user session and will dierenti-
ate even those users who share the same NAT IP address.
We showed that most of these connections can be learned if
the attacker maintains connections to a majority of Bitcoin
servers. Then we show that the transaction propagation
rules imply that the entry nodes will be among the rst
that report the transaction to the attacker. As soon as the
attacker receives the transaction from just 2-3 entry nodes
he can with very high probability link the transaction to a
specic client. Moreover a sequence of successfully mapped
transactions can help the attacker to track dynamic changes
in the entry node set, to keep the client identier fresh. The
cost of the deanonymisation attack on the full bitcoin net-
work is under 1500 EUR.
We demonstrate that the use of Tor does not rule out
the attack as Tor connections can be prohibited for the en-
tire network. Our technique is orthogonal to the transac-
tion graph de-anonymisation techniques and can be used in
combination with them. It shows that the level of network
anonymity provided by Bitcoin is quite low. Several features
of the Bitcoin protocol makes the attack possible. In partic-
ular, we emphasize that the stable set of only 8 entry nodes
is too small, as the majority of these nodes connections can
be captured by an attacker. A countermeasure could be to
randomize and regularly rotate these nodes, and to submit
transactions via another set of nodes.
We also described a number of techniques that reveal the
topology of the Bitcoin network. Some of them are used for
our attack, but the entire set is interesting by themselves
and not only in the context of deanonymisation. For exam-
ple it can be used as a tool to better understand relations
between bitcoin peers (e.g. one can check if peers of major
pools keep permanent connections between each other). As
another example, an adversary can nd the minimal cut in
the network graph and target those connections with denial-
of-service attacks (an example of a memory exhaustion at-
tack that we discovered while digging through the bitcoin
source code can be found in the appendix). This would re-
sult in splitting the network in two parts. Our results open
several directions for the future research.
Yet another feature is the lack of authentication within the
network, which requires the nodes to blacklist misbehaving
peers by IP. We gured out that very short messages may
cause a day IP ban, which can be used to separate a given
node or the entire network from anonymity services such as
proxy servers or Tor. If the Bitcoin community wishes to
use Tor, this part of the protocol must be reconsidered.
Finally, we showed that the routine procedure of adding
a checkpoint to the client code might be exploited to con-
struct an alternate reality. While too noticable as an attack
scenario, this idea can be a solution in the case of unforeseen
and unsustainable rise of diculty.
11. REFERENCES
[1] Bitcoin code project, v.0.8.6.
https://github.com/bitcoin/bitcoin, 2014.
[2] Bitcoin wiki. https://en.bitcoin.it/wiki/, 2014.
[3] Bitcoins in space. http://www.virgin.com/
richard-branson/bitcoins-in-space, 2014.
[4] Bitnodes. https://github.com/ayeowch/bitnodes,
2014.
[5] Onioncat an anonymous vpn-adapter.
https://www.onioncat.org/about-onioncat/, 2014.
[6] D. Chaum, A. Fiat, and M. Naor. Untraceable
electronic cash. In Proceedings on Advances in
Cryptology, CRYPTO 88, pages 319327, New York,
NY, USA, 1990. Springer-Verlag New York, Inc.
[7] Roger Dingledine, Nick Mathewson, and Paul
Syverson. Tor: The second-generation onion router.
Technical report, 2004. http:
//www.dtic.mil/dtic/tr/fulltext/u2/a465464.pdf.
[8] J. Laurie. Grinstead, Charles M.; Snell. Introduction
to Probability. American Mathematical Society, 1997.
[9] Philip Koshy, Diana Koshy, and Patrick McDaniel. An
analysis of anonymity in bitcoin using p2p network
trac. 2014. to appear at Financial Cryptography14.
[10] Sarah Meiklejohn, Marjori Pomarole, Grant Jordan,
Kirill Levchenko, Damon McCoy, Georey M. Voelker,
and Stefan Savage. A stful of bitcoins:
Characterizing payments among men with no names.
In Proceedings of the 2013 Conference on Internet
Measurement Conference, IMC 13, pages 127140,
New York, NY, USA, 2013. ACM.
[11] Satoshi Nakamoto. Bitcoin: A peer-to-peer electronic
cash system. 2009.
http://www.bitcoin.org/bitcoin.pdf.
[12] Fergal Reid and Martin Harrigan. An analysis of
anonymity in the bitcoin system. In Security and
Privacy in Social Networks, pages 197223. Springer,
2013.
[13] Dorit Ron and Adi Shamir. Quantitative analysis of
the full bitcoin transaction graph. In Financial
Cryptography and Data Security, pages 624. Springer,
2013.
APPENDIX
A. ESTIMATINGSUCCESS RATE: DETAILS
In this section we describe a mathematical model that
allows us to estimate the success rate of the deanonymization
attack.
As inputs, we take the average probability p
addr
over the
network, which is estimated in Section 7.1, and the distribu-
tion of the number of entry nodes among the rst 10 nodes
reporting a transaction to attackers peers (Section 6). We
extrapolate the latter probability spectrum from the test
net to the main net, which assumes similar network perfor-
mance and the stability of the spectrum when the attacker
has more or fewer connections to servers. The correctness
of the extrapolation can be tested only by mounting a full-
scale attack on the network, which we chose not to perform
for ethical reasons.
First, we introduce two combinatorial formulas. Suppose
that there are N balls. If each ball is red with probability
pa, and green with probability 1 pa, then the probability
that there are R red balls is
P1(R; N) =

N
R

p
j
a
(1 pa)
NR
(1)
Now assume that there are R red balls and N R green
balls. Suppose that we select L balls at random out of N.
The probability that there will be exactly q red balls among
L chosen is computed as follows:
P2(q ; L, R, N) =

R
q

NR
Lq

N
L
.
Now we get back to Bitcoin. If each entry node is detected
with probability p
Avg
addr
= 0.34 (Section 7.3), then according
to Eq. (1) we detect R entry nodes out of 8 with the following
probability spectrum:
P1(R; 8) :
Nodes Probability
1 0.15
2 0.27
3 0.28
4 0.18
5 0.07
6 0.02
7 0.002
8 0.0002
According to our experiments on the Bitcoin test net (Sec-
tion 6), the probability to have L entry nodes among the
top-10 is as follows:
P3(L) :
Nodes Probability
1 0.02
2 0.055
3 0.1225
4 0.245
5 0.2125
6 0.2125
7 0.0925
8 0
We assume that both events are independent. Then the
probability that at least M out of these L nodes we have
detected (i.e. it belongs to the set of R entry nodes) is
Psuccess(M) =

qM

L8

R8
P2(q ; L, R, 8) P1(R; 8) P3(L);
We have made some calculations and got the following re-
sults:

L8

R8
P2(q ; L, R, 8) P1(R; 8) P3(L) :
L Probability
1 0.366
2 0.243
3 0.9
4 0.02
5 0.002
Psuccess(M) :
M Probability
1 0.721
2 0.355
3 0.112
4 0.022
5 0.002
Therefore, we expect to catch 3-tuples in 11% of transac-
tions, and 2-tuples in 35% of transactions.
We applied this model to the testnet as well, and obtained
that it ts our actual deanonymization results well:
Estimated p
addr
Deanonymization rate with 3-tuples
Actual Predicted
0.64 41% 43%
0.86 59.9% 65.6%
B. ATTACK COSTS
The expenses for the attack include two main components:
(1) renting machines for connecting to bitcoin servers and
listening for INVENTORY messages; (2) periodically advertising
potential client addresses in the network. Note that if an
attacker rents servers, the incoming trac for the servers is
normally free of charge. Assuming that an attacker would
like to stay stealthy, she would want to have 50 dierent
IP addresses possibly from dierent subnetworks. Thus she
might want to rent 50 dierent servers. Assuming monthly
price per one server 25 EUR, this results in 1250 EUR per
month.
When advertising potential client addresses, the attacker
is interested in that the addresses propagate in the network
as fast as possible. In order to achieve this the attacker
might try to advertise the addresses to all servers simulta-
neously. Given that there are 100,000 potential clients and
the attacker needs to send 10 addresses per ADDR message,
this results in 10,000 ADDR messages of 325 bytes each per
bitcoin server or (given there are 8,000 bitcoin servers) 24.2
GB in total.
If an attacker advertises the addresses every 10 minutes
and she is interested in continuously deanonymising trans-
action during a month, it will require sending 104,544 GB
of data from 50 servers. Given that 10,000 GB per server is
included into the servers price and the price per additional
1,000 GB is 2 EUR, the attacker would need to pay 109 EUR
per month. As a result the total cost of the attack is esti-
mated to be less than 1500 EUR per month of continuous
deanonymisation.
C. TRANSACTIONPROPAGATIONDELAY
In this section we measure transaction propagation de-
lays between our high-speed server (1 Gbit/s, Intel Core i7
3GHz) located in Germany and 6,163 other bitcoin servers.
As was described in Section 2, it takes 3 steps to forward a
transaction between two bitcoin peers. As we are not able
to obtain times when a remote peer sends an INVENTORY mes-
sage, we skipped the rst step (i.e. propagation delays of
INVENTORY messages) and measured time dierences between
receptions of corresponding INVENTORY messages and recep-
tions of the transactions. Note however that the size of an
INVENTORY message is 37 bytes, while the size of a transac-
tion which transfers coins from one pseudonym to two other
pseudonyms is 258 bytes. Thus the obtained results can
serve as a good approximation. For each bitcoin server we
collected 70 transactions and combined them into a single
dataset (thus having 431,410 data points). Fig. 10 shows
probability density function of the transaction propagation
delay between our node and other bitcoin servers and Fig. 11
shows the corresponding cumulative distribution.
0
2
4
6
8
10
12
0 50 100 150 200 250 300 350 400 450
%

o
f

c
o
n
n
e
c
t
i
o
n
s
Time, seconds
Figure 10: Transaction propagation delay, density
0
20
40
60
80
100
0 200 400 600 800 1000
%

o
f

c
o
n
n
e
c
t
i
o
n
s
Time, seconds
Figure 11: Transaction propagation delay, cumula-
tive
D. DENIAL OF SERVICE
In this section we analyse the security of bitcoin network-
ing protocol against Denial of Services attacks.
D.1 Memory exhaustion by address ooding
Bitcoins peer discovery protocol has a mechanism which
prevents multiple retransmissions of the same addresses: for
each connection it has, a bitcoin node maintains a history
(Implemented as an instance of std::set C++ class) of ad-
dresses which were sent over this connection. This history
is emptied once per every 24 hours and more importantly
does not limit the number of elements it holds. In order to
check if one can ood this container with fake addresses we
conducted a simplied experiment. We set up locally two
bitcoin nodes so that when one of the nodes (the target) re-
ceives an ADDR message it forwards the addresses it contains
to just one neighbour. Both machines had Ubuntu 12.04
installed with 2Gb of RAM and the same amount of swap
memory. They were running bitcoind version 0.8.6.
We were sending fake addresses with the rate of 30,000
addresses per second. After approximately 45 minutes, the
response delay to the users interactions became signicant
and the node was unreachable for new bitcoin connections.
We also mounted a reduced version of this attack on our
own bitcoin node in the real network. We terminated the
experiment when the memory consumption increased by 100
MB.