computers & security 99 (2020) 102033

Available online at www.sciencedirect.com

journal homepage: www.elsevier.com/locate/cose

A trusted feature aggregator federated

learning for distributed malicious attack detection

Xinhong Hei, Xinyue Yin, Yichuan Wang∗, Ju Ren, Lei Zhu

Xi’an University of Technology School of Computer Science and Engineering, China

a r t i c l e i n f o a b s t r a c t

Article history: With the rapid development of IoT technology, millions of physical devices embedded with
Received 30 March 2020 electronics or software are put into regular production. Each IoT device is connected to the
Revised 31 August 2020 user’s life and property privacy. Without a credible intrusion detection and defense mech-
Accepted 2 September 2020 anism installed on the device, it may be attacked by hackers, such as monitoring events of
Available online 7 September 2020 home cameras and control of smart devices. These attack events will have a serious impact
on users’ production and life. This paper proposes a Blockchained-Federated Learning based
Keywords: cloud intrusion detection scheme. The scheme sends the local training parameters of the
IoT IoT intrusion alarm set to the cloud computing center for global prediction, and stores the
Cloud security model training process information and behavior on the blockchain. In order to solve the
Privacy protection high probability of false alerts affecting the accuracy of the federated learning model, the
Federated learning scheme proposes an alert filter identification module. At the same time, through the erasure
Blockchain code-based blockchain storage solution, the traditional blockchain storage performance is
improved to meet the storage needs of a large number of alert training data in real scenarios.

© 2020 Elsevier Ltd. All rights reserved.

attack methods in different application scenarios. The tradi-

1. Introduction tional network security mechanism cannot be fully or adapted
to the existing technology. In the absence of a reliable secu-
The Internet of Things(IoT) is the use of information collec-
rity mechanism, the more IoT devices penetrate into our lives,
tion equipment such as radio frequency identification, sen-
the more vulnerable users’ privacy and even the safety of life
sors, infrared sensors, laser scanners, etc., according to spe-
and property are. The information interaction in the current
cific protocols, to connect computable devices to the Inter-
IoT environment requires a security mechanism to solve the
net. Build an IoT environment for edge-side information ex-
problems of massiveness, sensitivity, siloing and security of
changes and communication to achieve intelligent identifica-
multi-party data computing.
tion, location, tracking, surveillance, and management in spe-
Regarding the complex security issues of IoT, IoT secu-
cific scenarios. In recent years, the IoT technology has been
rity researchers have proposed different security protection
widely used in smart city, wearable device (Ren et al., 2017),
strategies. Ren et al. (2016) propose a channel-aware reputa-
agriculture, manufacturing industry, etc., which has led to a
tion system with adaptive detection threshold, and find a so-
surge in the number of connected devices, which has reached
lution to the danger of wireless sensor networks being vul-
billions (Soldatos et al., 2015). As the scale of the IoT system
nerable to selective forwarding attacks. Wang et al. (2020)pro-
continues to expand, it also brings security challenges and a
pose a scheme for syntactic analysis of unknown protocols
crisis of privacy disclosure. The network security environment
based on convolutional neural networks. The scheme helps
cannot be predicted. The IoT devices are subject to special
to measure the boundary of a botnet in an IoT environment.

∗
Corresponding author.
E-mail address: [email protected] (Y. Wang).
https://doi.org/10.1016/j.cose.2020.102033
0167-4048/© 2020 Elsevier Ltd. All rights reserved.
2 computers & security 99 (2020) 102033

Intrusion detection system(IDS) is a commonly used security • Adopt Hyperledger Fabric expansion scheme based on era-
strategy in current work. Zarpelo et al. (2017) mentioned that sure codes. Reduce the storage pressure of the blockchain,
the IDS is one of the core components of the IoT system. The improve the storage performance of the blockchain, and
IDS is deployed in a heterogeneous IoT environment to en- meet the storage needs of a large number of alerts training
hance the security and robustness of the system. As a security data in real scenarios.
middleware, IDS has been widely deployed in IoT security ap-
plications. Meng and Kwok (2011)proposed an adaptive expert The rest of the paper is organized as follows:
knowledge-driven real-time IDS. At the same time, it is a com- Section 2 presents the background as well as the princi-
prehensive detection scheme for specific network functions. ple of the federated learning and blockchain,and the current
Midi et al. (2017) proposed a method to analyze abnormal net- research work; Section 3 describes the architecture and
work traffic by using the deep autoencoders method to extract workflow of the scheme; The alert filter identification module
network behavior, and perform intrusion detection and pro- and blockchain storage module are introduced in Section 4.
tection on the damaged IoT devices. Meidan et al. (2018) a real- Section 5 is experimental analysis. Section 6 carries on the
time hybrid network intrusion detection framework based on comprehensive summary of this scheme.
MapReduce architecture is used to prevent two common rout-
ing attacks—tiankeng attack and selective forwarding attack.
Li et al. (2020) proposed to collaborative IDS for active in- 2. Related work
trusion detection, combined with multiple IDSs for data anal-
ysis, to judge and predict the detection results. But collabo- In this section, we introduce the technical background related
rative IDS has the drawbacks of delay and no real-time. The to our system, such as federated learning and blockchain.
traditional cloud intrusion detection technology is that the Train alert samples of IoT devices through federated learn-
IoT device sends the local detection results to the cloud for ing, while building trusted execution environments with
analysis and judgment, but the traditional solution has some blockchains for federated learning. At the same time, the re-
drawbacks. Firstly,The detection result set will expose the vul- search results of Privacy-Preserving Machine Learning (PPML)
nerable port of the damaged IoT device. If the detection set is are discussed in the federated learning section. Finally, the re-
maliciously captured by an attacker, it will cause a greater se- search achievement of the combination of federated learning
curity crisis. At the same time, the detection data set is also and blockchain technology are analyzed.
the sensitive data of the device. If the device owner refuses to
send it to the cloud, it will lead to data islands and the cloud 2.1. Federated learning
detection library cannot be updated. Furthermore, there are a
lot of false alarms in the detection alarm set, which increases Federated learning (FL) is a training and analysis of local
the workload of the cloud server. data of the data owner with multiple participants under the
This paper proposes a Blockchained Federated Learning premise of protecting sensitive data. FL is a distributed ma-
cloud intrusion detection system(BFL-CIDS) scheme. Under chine learning architecture. FL consists of central server G and
the premise of protecting the sensitive detection alert infor- multiple local devices {C}N i=1 . In epoch t, there are G partici-
mation, the distributed machine learning trains the local data pants PK , Dk represents the data set owned by the k-th par-
set to generate weight parameters, and passes the weight pa- ticipant. The local device trains the global model MtG and the
rameters for global learning and prediction. However, IoT de- local data set DK through a random descent gradient (SGD)
vices have weaknesses in storage capacity and computing algorithm. Generate local parameter weight wtk . Multiple par-
power (Wang et al., 2019), and cannot perform local training ticipants send this round of parameter weights to the central

better. At the same time, uploading a large number of weight server G, G set N t
k=1 wk through the federated average algo-
parameters to the server will also cause broadband conges- rithm training to form global parameter weights WGt and new
tion, which is also one of the disadvantages of federated learn- model MG (t+1) . Let the ideal model for training be M, the loss
ing. For the low storage capacity and computing power of iot function ∇L( · ), and the learning rate λ. mi (wtk ) = λ∇L(wtk , Dk ).
devices,we propose to set up regional service party to collect NK = |PK | indicates the cardinality of PK .
the detection and alert sets of IoT devices in the region. The re-
gional service party performs local training after filtering the 
K
Nk 1 
M= M (wt ) Mk (wtk ) = mi (wtk ) (1)
false alert information, and stores all the training results on N k k Nk
k=1 i∈Pk
the blockchain. It ensures that the training results are perma-
nently saved and cannot be tampered with. This scheme pro- If the data set DK participating in federated learning is in-
vides CIDS with functions of data privacy security, preventive dependently identically distribution(IID), you can get expecta-
maintenance, and permanent storage. tions EDk [Mk (wtk )] = M. But in more practical applications, the
The research contributions of this article are: data set is Non-IID (Zhu et al., 2018), which is also one of the
problems to be solved by federated learning. If DK is Non-IID,
the global model wk ( · ) maintained by Pk is very close to the
• Provide privacy protection for CIDS through Blockchained target model.
federated learning; As data breaches and privacy infringement events are
• Perform false alerts filtering on the detection alert set, re- well known by the public, privacy-oriented machine learn-
duce the working pressure in the cloud, and improve the ing (PPML) has become a security solution that focuses on.
quality of the federated learning model. Possible adversaries of FL are semi-honest adversaries and
 computers & security 99 (2020) 102033 3

malicious adversaries. As for the way to protect data, exist- When a node initiates a transaction, it is necessary to use
ing research mainly focuses on secure multi-party comput- a private key for digital signature. Hyperledger Fabric uses
ing (Lindell, 2006), homomorphic encryption (Xie et al., 2014), an Elliptic Curve Digital Signature Algorithm (ECDSA) as
and differential privacy Abadi et al. (2016). (Mohassel and digital signature algorithm. ECDSA is one of the application
Zhang, 2017) is a PPML two-party (client, server) framework examples of elliptic curve cryptography (ECC). ECC can use a
that uses a two-stage training model. Use linear homomor- shorter key than the asymmetric encryption algorithm RSA
phic encryption or inadvertent transmission to generate the to achieve the same strength.
necessary triplets to complete the offline training task. Of- Elliptic Curve Digital Signature Algorithm (ECDSA): Let the
fline tasks mainly serve multiplication operations in the on- private key and public key be k and K respectively. Where K =
line phase. For the nonlinear activation function, an alterna- kG and G is a point on the elliptic curve E. A random number
tive solution to support safe multi-party calculations is pro- r is randomly generated before encrypting the plaintext M.
posed. Bostani and Sheikhan (2017) combining FL with multi- Private Key Signature:
party secure computing (MPC), a protocol is proposed for the
(h+kx )
secure aggregation of high-dimensional data. The agreement φ = rG(x, y ); h = Hash(M ); s=
  r (2)
is applicable to large-scale terminals to complete the input Commit Peer ← M, (φ, s )
sum through the same server. But the premise is that specific
terminal input cannot be exposed. Chai et al. (2019) realized Peer Public Key Verification Signature:
and improved the SPDZ2k based on the active safety multi- hG+xK
Compute h = Hash(M ); ψ= s
party computing protocol proposed by Cramer et al. (2018), and (3)
Veri f y φ == ψ
realized the inadvertent evaluation algorithm of decision tree
and support vector machine (SVM). Chai et al. (2019) through The key to ECDSA is the introduction of a random num-
the Paillier homomorphic encryption algorithm, the safe fed- ber r, which improves the security of the signature. Even if the
eration matrix factors under the conditions of honest cus- same message is changed, as long as the random number r is
tomers and honest but curious servers are decomposed. changed, the resulting signature will change accordingly.
Geyer et al. (2017) proposed a Differentially Private Federated
Learning model to prevent the training model from being sub- 2.3. Research works of blockchained-FL and IoT
jected to differential attacks. Lu et al. (2020) integrate local dif-
ference privacy into FL to protect the privacy of the updated The traditional distributed federated learning system pro-
local model. Establish a secure federated learning program tects the user’s privacy by training the user’s data locally and
through differential privacy. uploading the training parameters to the central computing
node. However, at present, the federated learning is not per-
2.2. Hyperledger fabric fect, and the user privacy data can be obtained by training
parameter backward reasoning, and the process of parame-
Hyperledger Fabric is a permissioned blockchain ter transmission is not credible, which cannot achieve the real
(Androulaki et al., 2018) that must be identified and approved comprehensive protection of user data (Ma et al., 2019). Im-
by the organization before accessing the Fabric network. All provements can be made in the storage, transmission, and
members jointly maintain a distributed ledger. All data opera- traceability of training parameters and models. The BFL-CIDS
tions are public and will be permanently recorded and cannot scheme creates a trusted execution environment for federated
be tampered with. The identity of the node in Fabric is real- learning by combining federated learning with permissioned
ized by digital certificate. Digital certificates are managed and blockchain. In order to prevent data tampering, the security
issued by Fabric CA. Chaincode is an electronic code that is not of training parameters is guaranteed and reliable data sup-
controlled by a third party and will be automatically executed port is provided for the correct model of training. At the same
when the condition is triggered. At the same time, Chaincode time, blockchain provides traceable data tracking for federated
can meet different business needs and realize the application learning and encryption and protection for training parame-
of blockchain in different scenarios. Hyperledger Fabric’s ters.
innovations in modules such as multi-channel, multi-ledger, At present, there are also many research works on
rights management, and consensus mechanisms provide the combination of federated learning and blockchain.
new ideas for the development of executable services on the Martinez et al. (2019) proposed to use the EOS blockchain as an
blockchain. Fabric’s private data can be isolated by unique incentive layer for federated learning. Through records and in-
organizations and channels, providing more data protection centives, the enthusiasm and high-quality data contributions
methods (Benhamouda et al., 2018). In the Hyperledger Fabric of federated learning participants are guaranteed. Majeed and
network, all peer nodes may have three identities (can have Hong (2019) designed the FLchain architecture and formed
more than two identities at the same time), namely Endorse a blockchain network by edge devices. Through the con-
Peer, Commit Peer and Orderer Peer. After peer initiates a cept of the blockchain channel, the global model is specif-
transaction, the client will first send it to the endorse peer ically allocated and the model is stored on the blockchain.
to simulate the transaction, then the transaction enters the Zhang et al. (2020) proposed a secure data sharing architec-
transaction pool. The orderer peer sorts block, and distributes ture based on the blockchain authorization of the Internet of
to the commit peer and the world state database for con- Vehicles to protect the privacy of shared data through fed-
firmation, after meeting the conditions of the consensus erated learning. Improve the utilization rate of system com-
mechanism. After that, the block data was officially launched. puting resources. Wang (2019) in view of gradient information
4 computers & security 99 (2020) 102033

Fig. 1 – BFL-CIDS scheme architecture.

leakage and vulnerability to integrity attacks, a blockchained aster tolerance The FL Serving module is an online inference
federated machine learning (BlockFedML) is proposed. At the to local model. FL Flow is a key module for executing task
same time, a security parameter aggregation mechanism, a scheduling, in which the client’s main task is the user sub-
checkpoint-only contract and an incentive mechanism were mission training task, the server is used to handle the user’s
developed. Yin et al. (2020) apply blockchain and FL to the submission request scheduling, and the server is the entrance
IoT environment to implement a secure data collaboration of the FL cluster. Federation is a module for learning data ex-
framework. Separate public data from private data, and pro- change between parties. FL Board is a visualization module
tect the safe use and transmission of data through blockchain. that can timely display the status, amount completed, train-
Sharma et al. (2020) proposed a distributed computing de- ing results and index values of tasks. The Roll module pro-
fense framework for a sustainable society, and proposed an vides distributed computing and storage support. Proxy mod-
algorithm to meet the challenges of limited training data. ule realizes model forwarding between parties and is the only
Ramanan and Nakayama (2019) proposed an FL environment network exit of parties, which can increase security and facil-
without aggregators, which eliminated the task of centralized itate network communication management. The parameters
aggregators through the blockchain. The efficient combination and models after the FL training are stored in the blockchain
of blockchain and IoT technology (Tian et al., 2020) proposed through the chaincode.
blockchain-based secure key management scheme to protect Chaincode Layer is a digital protocol that can be used to
the security of data collected by dynamic wireless sensor net- implement different blockchain application scenarios. Chain-
works(DWSNs) in the Industrial Internet of Things(IIoT), and code is not controlled by third-party members, and execution
improve the performance of DWSNs in IIoT Credibility. can be triggered when the protocol conditions are met. Chain-
code is divided into business chaincode and system chain-
code. The former is responsible for implementing specific
3. Scheme architecture and structure business rules, and the latter sets the system specifications for
Peer endorsement and transactions, block queries, and chain-
3.1. System architecture code lifecycle. This solution designs two types of business
chaincode, one is to upload weight parameter settings for re-
The architecture of BFL-CIDS scheme is shown in Fig. 1. There gional service party, and the other is to upload global train-
are four components including: Application layer, Federated- ing model settings for global training party. Through different
Learning layer, Chaincode layer,and Blockchain layer. The fol- business chaincode to achieve different role operation rules.
lowing is a detailed description of each layer. Blockchain Layer is based on open source Hyperledger Fab-
Application Layer is a collection of edge IoT devices. Each ric as the underlying blockchain technology. Chaincode layer
device is equipped with IDS to detect and monitor network interacts with the blockchain layer through gRPC. gRPC is a
attacks through IDS. Collect and send abnormal alarms to the modern open source high performance remote procedure call
federated learning layer. At the same time receive the system (RPC) framework that can run in any environment. In this so-
security patch sent by the federated learning layer. For the lution, the blockchain provides storage of training parameters,
latest abnormal network detection library, to avoid the same evidence chain, and traceability services. By permitting the
false positive alert multiple times. advantages of blockchain identity management, the operating
Federated Learning Layer. The FL framework can be divided environment of federated learning is purified and the quality
into offline training and online training, so as to ensure the of the training model is optimized.
high availability of the system and consider the system dis-
 computers & security 99 (2020) 102033 5

3.2. Scheme party description and stores it in a local database. In the Training module, RSP
performs local training according to the global model MtG , and
Learning − unabled Node(LN ) : LN is an IoT edge device with sends the weight parameter Wkt to GTP. In the Reporting mod-
weak learning ability and storage capacity. LN joins the BFL- ule, GTP aggregate participates in the training of RSP weight
CIDS scheme to monitor the abnormal behavior of the device parameters. GTP uses the FedAvg algorithm for global model
network. The caught exception behavior is sent to the RSP in training and obtains the new model Mt+1 t+1
G . Send MG to the RSP
the form of a warning log. After receiving the result of RSP node. In Training and Reporting module, RSP and GTP operate
feedback alert information, LN will add false positive alert to in the blockchain network. All training parameters and mod-
the white list to avoid false positive alert again, and timely up- els will be permanently saved to the blockchain. Each round
date the security patch of the defense system to enhance the of models and parameters can realize the functions of perma-
defense function of the system. The common abnormal alerts nent storage, non-tampering and real-time traceability.
are: abnormal process behavior, sensitive file tampering, ab- Fig. 3 is the network topology of the BFL-CIDS scheme. RSP
normal network connection, abnormal flow alert, application upon receive alerts submitted from LN by a region,RSP ini-
intrusion, etc. tially filters and classifies the alerts to store the available alerts
Regional Service Party(RSP) : RSP is the Service node that in local database. After RSP conducts local training on the
connects LN to GTP and is also a member node on the per- sample, the training parameters are formed and stored in the
missioned blockchain. RSP has certain computing and storage blockchain. GTP traces the latest training parameters of each
capacity, and can perform preliminary filtering and training RSP from the blockchain for global model learning. Based on
for the submitted alerts information. The perliminary filtering a large number of alert samples, the attack is analyzed and
work includes labeled false poitive alters and alerts classifica- predicted, the attack behavior or trend is inferred, and a new
tion. After local training, RSP send the common alerts and ab- model is trained to be uploaded to the blockchain. RSP and
normal alerts of LN in the region to GTP for centralized train- GTP, as the nodes of premissioned blockchain, jointly main-
ing in the form of model parameters. RSP will also upload the tain the training parameters and models.
model parameters to the permissioned blockchain to prevent
the parameters from being tampered in the transmission pro-
cess and ensure the reliability of parameters. The data on the 4. Detailed description of scheme
permissioned blockchain is maintained by all the members to-
gether, and no node can change the ledger information, pro- 4.1. Alert filter identification module
viding a secure and permanent storage for parameters.
Global Training Party(GTP) : GTP is the centralized Training 4.1.1. Alert analysis
node of the BFL-CIDS scheme and a member node in the per- The current memory capacity and computing capacity of edge
missioned blockchain. GTP has strong computing and stor- nodes of the IoT are not enough to meet the needs of mod-
age capacity, and aggerates all RSP submission parameters ern mass data. Each edge device has no memory for the tens
for global training. Based on a large number of attacks and of thousands of alerts it generates each day. The amount of
defense samples, after machine learning of alerts, the alerts alarm information for each node in the RSP collection area is
attack behavior is simulated to form a new training model. even larger. Worse, false alerts account for 90% of the unfil-
The model is divided into three types: the trend of alert at- tered alerts, and such a high false alert rate can reduce the ac-
tack, identifying new attack algorithm and non-threat alerts. curacy of federated learning models. Snort is an open source
GTP also uploads the model to the permissioned blockchain. intrusion detection system.We replay the data set DARPA1999
The RSP, as the node of the permissioned blockchain, will in- on the Snort system to test for false positive alerts. There
directly receive the global training results at the first time were no attacks on the first and third weeks of DARPA1999,
and feedback the final alarm status to LN. GTP can obtain the but Snort showed more than 6,000 alerts each week, which
previously uploaded training model from the blockchain, and were false alerts. The sheer volume of false postive alerts is
timely train the new model according to the latest submitted a huge challenge for federated learning.Therefore, RSP added
warning sample, in order to cope with the constant evolution the alert filter identification (AFI) module to preliminarily de-
of network attack means and virus samples. permissioned termine and identify alerts. Add the available alerts to the
blockchain provides a safe and permanent version traceabil- training database for machine learning. AFI modules can en-
ity model for the training model, and improves the application hance the performance of federated learning, reduce redun-
scenario and credibility guarantee of federated learning. dant training and ensure the quality of training models.
The BFL-CIDS solution replayed a low-rate DDoS attack
3.3. Scheme workflow and network model data set (The DDOS attack occurred at 16:42-17:00 on Decem-
ber 10, 2018) on the iot device to simulate the alarm handling
Fig. 2 shows the workflow of the BFL-CIDS scheme. In round of the iot device in the face of intrusion process. The Iot device
n, it is mainly divided into Collection module, Training mod- sends the abnormal behavior flow information process to the
ule and Reporting module. In the Collection module, RSP is RSP. After RSP sorts out the collected alarm traffic informa-
mainly responsible for the collection and filtering services of tion, the preliminary alarm information is obtained in table 1.
alerts. LN sends the alarm information to RSP, and RSP re- Perform the Apriori algorithm on the alarm information to
sponds to LN and updates the latest security patch. When the complete the frequent item set mining, and obtain the fre-
network connection fails or the device is damaged, the RSP re- quent item-level rules of the alarm traffic information. Based
quest will be rejected. The RSP filters the collected information on the prior knowledge (determined warning behavior) pro-
6 computers & security 99 (2020) 102033

Fig. 2 – The workflow of BFL-CIDS scheme.

the local rule base from the model parameters fed back by GTP
to obtain more accurate alarm judgments.
The solution can design different rules according to the fa-
teures of the traffic in formula (4). According to the abnormal
points of the intrusion information feedback, the relevant ab-
normal fateures are extracted. For example, if the packet flow
is abnormal, the fateures of packet_threshold are extracted.

[service : Protocoltype & Dstip & TimeStamps

(4)
& Duration & Features]{support & con f idence}

Design frequent item rule format for alarm information in

Table 1. Such as formula (5):

Fig. 3 – The network topology of BFL-CIDS scheme. [service : htt p & Dstip = 10.42.0.1
& TimeStamps = 2018 − 12 − 10; 16 : 42 : 18.5
(5)
& duration = 1061.1s & packet _t hreshol d = 4354]
Table 1 – Analyze alarm traffic information. {support = 5% & con f idence = 95%}

Srcip Dstip packats Duration The rule shows: When using http service to connect, for
10.42.0.215 52.40.109.206 30 16:43:42.4 the rule item whose destination ip(10.42.0.1) starts 16:42:18.5
-16:44:54.5 on December 10, 2018.The duration is 1061.1s and the packet-
10.42.0.29 10.226.11.155 4354 16:48:19.3 threshold is 4354. The support and reliability of this rule is
-16:49:37.8 respectively 5% and 95%.
10.42.0.106 - 10.42.0.1 2690 16:42:18.5
110.42.0.107(etc.) -16:59:59.6
4.1.2. Detailed description of AFI module
10.42.0.1 10.42.0.107 431 16:42:19.7
At present, the most common false alarms are characterized
-10.42.0.107(etc.) -16:59:35.4
10.42.0.1, 10.42.0.106 3855 16:42:18.5 by periodicity and high density. Extract the features of an alert
10.42.0.36 -16:59:59.6 as a set A:

A = {t imest amp, al ert _Type, prot ocal , src_IP,

(6)
dst _IP, dst _ port }
vided by GTP, RSP first marks some rules. RSP advances the rel-
According to Viinikka et al. (2006), false alerts are periodic
evant features of the labeled and unlabeled alarms, and then
and intensive. We define the two characteristics of a false alert
executes local training (a semi-supervised learning algorithm
as follows.
based on divergence). After multiple iterations of training, de-
termine whether the attribute of the alarm rule is a false alarm Definition 1. Alerts Periodicity Feature. According to the
or a true alarm. RSP will also timely update the alarm rules of statistics of alerts over a period of time at a specific time, the
 computers & security 99 (2020) 102033 7

Table 2 – Similar alert analysis.

Similar characteristics Description

Ai .al ert _Type==A j .al ert _Type & Ai .src_IP==A j .src_IP & Ai .dstIP ==A j .dst _IP The same security incident
Ai .src_IP==A j .src_IP & Ai .dstIP ==A j .dst _IP Same attackers against the same target IP
Ai .src_IP==A j .src_IP Multiple attackers against the same target IP
Ai .dst _IP==A j .dst _IP Same attackers against multiple target IPs

random process μ(t): t ∈ N is used to represent the generated In Algorithm 1, Ck represents k candidate item set,Fk stands
time series, and μ(t) is the number of alarms generated in time
T. Periodic analysis is carried out on the time series. The time
domain signal is transformed into frequency domain signal Algorithm 1 Frequent itemset generation of the Apriori algo-
by Discrete Fourier TransformDFT, and the accurate periodic rithm
value T is calculated. The definition of DFT is: 1: k=1.
2: Fk = {i | i ∈ I ∩ σ (i ) ≥ N ∗ minsup}.

N−1
t2π j f 3: repeat
DF T ( f ) = μ(t)e N (7)
n=0
4: k = k + 1.
5: Ck = apriorigen (F( k − 1)).
2π j
∈ C, C is the set of constant terms. Find the maximum 6: for each transaction t ∈ T do
N
value of f, fmax , and compute interval value T = f 1 . 7: Ct = subset (Ck , t );
max
8: for each candidate itemset c ∈ Ct do
Definition 2. If Alert A and Ā can satisfy one of the conditions 9: σ (c ) = σ (c ) + 1;
in Table 2, they are similar alerts. 10: end for
11: end for
Definition 3. Alerts Density Feature.Within the time interval
12: Fk = {c | c ∈ Ck ∩ σ (c ) ≥ N ∗ minsup}.
T, alert density Di is the ratio of the number of similar alarms
13: until Fk = ∅.
to the difference of their time distribution.
14: Result = ∪Fk .

Number(A,Ā )
Di = timax −timin . |timax − timin | < T
0, Otherwise.
for frequent k item set. N is the number of term sets. T is the
The AFI module uses frequent itemset mining algorithm number of transactions.subset() is defined as a function that
(Hei et al., 2019) and disagreement-based semi-supervised obtains subsets of candidates. apriori_gen() is the algorithm for
Learning (Tri-training algorithm) and the functions that the generating frequent k-item sets. minsup is minimum support
algorithm needs to achieve are as follows. Frequent itemset threshold. In steps 1-2, the algorithm first traverses the data
mining algorithm discovers redundant alert items in the alert set, detects the support degree of each item, and obtains fre-
log after the format is standardized. Judging the periodic item quent 1-item sets. Next in step 5, loop through the frequent
set and the alarm density according to the redundant alerts (k − 1) -itemset to derive k-candidate itemset. Calculate the
set, the false alarm is initially marked. The Tri-training al- candidate item set support by traversing the data set in steps
gorithm trains the marked and unmarked alerts in the local 6-10. After the support is calculated in step 12, the infrequent
database through three classifiers and filters false alerts. item set is eliminated. At step 13, the algorithm ends when no
Frequent itemset mining algorithm(Apriori algorithm):For new frequent item sets are generated.
I = {i1 , i2 , ..., id } is the set of all items in the data, while T = If we solve all the frequent item sets directly, the time com-
{t1 , t2 , ..., td } is the set of all transactions. A collection of 0 or plexity will be very high. Apriori algorithm can reduce the time
more items is called an itemset. If an item set contains k items, complexity of frequent item sets. The time complexity of Apri-
it is called an k-item set. Obviously, each transaction n con- ori algorithm is related to minimum support threshold and
tains a subset of I. transaction number. Because frequent itemset generation of
The association rule is the implication expression of X ← Y, the Apriori algorithm is nested with this apriori_gen method,
where X and Y are disjoint item sets, namely X ∩ Y = ∅. The the time complexity of Algorithm 1 is O(n2 ).
strength of an association rule can be measured by its Support Disagreement-based Semi-supervised Learning was pro-
and Confidence. The support determination rule can be used posed by Blum and Mitchell (1998). They assume that the
for the frequency of a given data set, while the confidence de- dataset has two fully redundant views, each view has a strong
termines the frequency of Y in transactions involving X. The learner, and two views Mutually independent under condi-
forms of Support s and Confidence c. are defined as follows: tions. Zhou and Li (2005) proposed a single-view, tri-training
algorithm (Algorithm 2). A significant feature of the algorithm
σ (X ∪ Y )
s(X ← Y ) = is that it uses three classifiers, which can not only easily han-
N
dle the problem of label confidence estimation and prediction
σ (X ∪ Y )
c(X ← Y ) = (8) of unseen examples. The flow of the algorithm is mainly as
σ (X)
follows:
8 computers & security 99 (2020) 102033

Algorithm 2 Tri-training algorithm

Input: L:Labeled training set; U:Unlabeled training set;
LA:Learning Algorithm
1: for i ∈ 1,2,3 do
2: Si ← Boot st rapSample(L )
3: hi ← LA(Si )
4: ei ← 0. 5;li ← 0
5: end for
6: repeat until None of hi (i = 1, 2, 3 ) changes
7: for i ∈ 1,2,3 do
8: hi ← ∅; U pdatei ← false
9: ei ← MeasureError(hm , hs )(m, s = i ).
10: if (ei < ei ) then
11: for Each x ∈ U do
Fig. 4 – Alerts filter and identification model.
12: if hm (x ) = hs (x )(m, s = i ) then
13: Li ← LiU (x, hm (x ) )
14: end if
15: end for alarm sample set and converted into a common format. Fre-
16: end if quent feature itemset mining component is applied to the
17: end for data items to obtain similar alerts information and conduct
 
18: if li = 0 then the first information filtering and analysis. The third compo-
ei
19: li ← e −e +1 nent is to save the data items to the database for backup and
i i
20: end if read the labeled and unlabeled data items from the database.
 
21: if li < |Li | then Disagreement-based semi-supervised learning component is
 
22: if ei |Li | < ei li then through three different classifiers for the marked and the un-
23: U pdat ei ← t rue marked items of data through the tri-training algorithm are
e
24: else if li > i i then analyzed. Filtration component on the step analysis to filter
ei −ei
e l results and true alerts output.
25: Li ← Subsample(Li , ie i − 1), U pdatei ← true
i
26: end if
27: end if 4.2. Blockchain storage based on erasure code
28: for i ∈ 1,2,3 do
29: if U pdat ei = t rue then The current blockchain storage mechanism uses multi-copy
30: hi ← LA(LULi ); ei ← ei ; li ← |Li | redundancy, and the nodes participating in the consensus in
31: end if the blockchain network need to store an identical copy.When
32: end for the blockchain needs to store too much data, it increases

33: Output: f (x ) ← argmaxy∈label h (x )=y 1 the burden of nodes. Lu et al. (2019) also mentioned that the
i
blockchain technology will enable federated learning, and the
system will be applied to the Internet of vehicles environ-
1) Repeatable sampling (bootstrap sampling) of the labeled ment. Faced with a large amount of system log information,
example set to obtain three labeled training sets, and then Lu chooses to upload node behavior to the blockchain. This
generate a classifier from each training set. paper proposes a blockchain storage model based on erasure
2) In the collaborative training process, the new labeled ex- code and discusses the advantages of this method over redun-
amples obtained by each classifier are provided by the dant copy storage in storage occupancy and system availabil-
other two classifiers. Specifically, if the two classifiers pre- ity.
dict the same unlabeled example, the example will be It is Erasure correction code was first used in the communica-
considered to have higher label confidence and is added to tion industry, and then it was widely developed in distributed
the labeled training set of the third classifier after labeling. storage. Reed - Solomon (RS) is by far the most popular remedy
3) Predict unlabeled instances through the integrated voting delete code, is a kind of support arbitrary code number n, and
of three classifiers. check the number m every block of Maximun short separable
(MDS) coding. RS code performs polynomial calculations for
Based on the above characteristics of error warning, coding segments and check segments in Galois field GF (2m ).
we propose an RSP alert filter identification(RSP-AFI) algo- Thai et al. (2015) indicates that when m:=8, GF has the best
rithm,which based on frequent item mining algorithm and computational efficiency.
semi-supervised learning algorithm. Fig. 4 shows the alerts fil- In the Hyperledger Fabric, the blockfile BF of j-th block is
j j j
ter Identification model. The scheme mainly consists of five segmented into m blockfile F0 , F1 , ..., Fm−1 . Encode m block-
components:alerts standardization,frequent feature itemset files under finite field GF to generate check blocks r =
mining,alerts storage, disagreement-based semi-supervised {ξ0 , ξ1 , ..., ξr }.Specifies the maximum memory Sm for the block-
learning and filtration. The first component aims to standard- file.Before sharding the blockfile, fill the source blockfile with 0
ize the alert information.the features are extracted from the to reach the maximum capacity Sm .This causes each blockfile
 computers & security 99 (2020) 102033 9

to be split evenly Smm bytes. The distributed storage method

Table 3 – Experimental configuration.
is adopted to store m blockfiles and r check blocks on dif-
ferent blockchain nodes respectively, so as to ensure that all Environment Version Notes
data can be restored when some nodes are offline or used in Operating system Ubuntu 18.04 -
the semi-honest user model environment. Adding RS code to Blockchain Hyperledger Premissioned
blockchain storage technology can reduce the storage com- Network Fabric 1.4.4 blockchain
plexity from O(n) to O(1). The GF(2m ) domain is a scalar multi- Go v1.13 Development
language
plication. A jth blockfile encoding is defined in the Eq. (9), and
Docker v19.03.1 Open source
set the invertible matrix vector be α. container
j j j j j j
Snort v2.9.9.0 Open source IDS
ξr j = αr·0 F0 + ... + α(r−1 )·(m−1 ) Fm−1 + αr·m Fm (9) hikey960 CPU:4 Cortex A73 Development Board

Assume that each organization on Hyperledger Fabric has

only one node. There are n nodes on the blockchain, sumB
blocks. Through RS code, BF is divided into m blockfiles and True Postive (TP): The instance is positive and is predicted to
r check blocks (n ≥ (m + r ), m > r). Storage optimization anal- be positive;
ysis of Fabric storage capacity expansion scheme. Blockchain False Negative (FN) : The instance is positive but is predicted
storage capacity SEC based on erasure code is: to be negative;
Sm m+r False Postive (FP): The instance is negative but is predicted
SEC = · numB · (10) to be positive;
m n
True Negative (TN): The instance is negative but is predicted
Traditional blockchain storage capacity S is: to be negative.
For each obfusion matrix, there is a set of metrics to eval-
S = numB · Sm · n (11)
uate. The following four evaluation methods are also the cal-
culation standards of the model adopted in this experiment.
The storage ratio ρ of SEC and S is:
Accuracy: The proportion of the result judged correctly in
m S ·num ·(m+r ) the total observed value as a result of the classification model.
SEC B
n·m m+r
ρ= = = (12) The calculation formula is as (13).
S numB · Sm · n m · n2

When n = m + r, ρ = m1·n . The optimization effect depends TP + TN

Accuracy = (13)
on the number of m and n, which is a considerable blockchain TP + TN + FP + FN
expansion plan.
Precision: In all results where the model predicts a positive
When the number of encoded partitions (m + r ) accessible
class, the model predicts the correct proportion. The calcula-
in the Fabric is less than m, the changed blockfiles cannot be
tion formula is as (14).
recovered. That is, when all the nodes storing blockfiles are
offline or the number of offline organizations is greater than or
TP
equal to mr·n+r , an unrecoverable situation may occur. But this Precision = (14)
TP + FP
is a kind of extreme condition. As a permission blockchain, the
control of identity authentication is more strict. Therefore, the Recall: The model predicts the correct weight among all re-
fabric expansion scheme based on erasure codes can ensure sults whose true value is positive. The calculation formula is
the availability of blockfiles. as (15).

TP
5. Experimental configuration and analysis Recall = (15)
TP + FN

5.1. Experimental environment and evaluation indicators F 1 − Score: Combines the results of Precision and Recall.
The calculation formula is as (16).
In the experimental environment of this paper, the operating
system is Ubuntu, and the blockchain project is based on the Precision ∗ Recall
F 1 − Score = 2 · (16)
hyperledger fabric. In the experiment, four hosts in the same Precision + Recall
LAN are connected to a blockchain network to upload and re-
trive the experimental data. The experimental confguration is 5.2. Experimental design and analysis
shown in Table. 3.
Model evaluation indicators. Obfusion matrix is often used <Member rule Settings> In the Hyperledger Fabric experiment,
to characterize the classification accuracy program of a classi- each node is assigned an organization(Org) to facilitate the in-
fier. In Rahman et al. (2020), the fractional calculation method dependence of the business. Two channels are set, RSchan-
of obfusion matrix is used to evaluate the model. For bi- nel and GTchannel. The three experiment nodes will be set as
nary confounding matrices, the predicted results can be di- RSP to be allocated to Org1, Org2, and Org3, and set one ex-
vided into four categories (TP, FN,FP and TN). The classification periment node as GTP to be allocated to Org4. According to
terms are described below. the business requirements of parties, Org1, Org2 and Org3 of
10 computers & security 99 (2020) 102033

Fig. 6 – The time of 1000 transactions between Ethereum

and Fabric.
Fig. 5 – the comparison of uploading time between
Ethereum and Fabric.

three RSP organizations upload model parameters in RSchan-

nel through RSchaincode. The Org4 of GTP organizations up-
loads global model in GTchannel through GTchaincode. The
separate channel is a blockchain ledger, which is separated by
different channels, so that the training parameters and train-
ing model can have independent ledger database, and the re-
trieval efficiency of the ledger can be improved. When all ver-
sions of the global model need to be extracted, the entire ac-
count of GTchannel can be traced locally to facilitate direct
extraction of the entire training model. When data retrieval is
performed according to GetHistoryForKey GHFK(k), all states of
k from genesis block to the last block of the given range are
scanned (Gupta et al., 2018). If store both RSP and GTP busi-
ness in the same ledger,will be to increase the time of data
retrieval and tracing, and cause tedious work for separating
training parameters and training model.

5.3. Analysis of experimental results

The experimental part analyzes and discusses the BFL-CIDS Fig. 7 – The optimal blockfile size for processing data.
scheme from three angles. The first part is the time compar-
ison of uploading data between public blockchain Ethereum
and permissioned blockchain Fabric. The more efficient ex-
ecution rate of the permissioned blockchain is the reason
chosen by this scheme. The second part is the discussion of eration training with the graph convolutional neural network
the optimal block file segmentation size in the Fabric storage (Yao et al., 2019) and DNN (Vinayakumar et al., 2019) scheme
scheme based on erasure codes. By comparing the fixed file models.
Sm
m block storage time, the minimum time for uploading and Ethereum (Zhu et al., 2018) is a public blockchain that can
retrieving the fabric that meets the actual application is se- also perform specific business functions through smart con-
lected. The third part compares the alarm recognition classi- tracts. Upload sample data of different sizes to Ethereum and
fication scores of RSP-AFI algorithm and several popular ma- Hyperledger Fabric to obtain the upload time of samples of dif-
chine learning algorithms in KDDCup99 data set. The alarm ferent sizes. Fig. 5 is the comparison diagram of sample upload
classification accuracy of the filtering algorithm is tested by time in Ethereum and Fabric.Experimental data show that the
the recognition rate of different identifiers. And the compar- time to store data on permissioned blockchain is better than
ison results of AUC value of the proposed model after 100 it- that in public blockchain.
 computers & security 99 (2020) 102033 11

In the same experimental environment, 1000 transactions

are uploaded to Ethereum and Fabric. As shown in Fig. 6,
the time comparison of 1000 transactions under different
blockchain is shown. The total elapsed time of 1000 trans-
actions uploaded in Ethereum is much greater than that of
Fabric. In some specific application scenarios, permissioned
blockchain has advantages over public blockchain in terms of
security and performance.
In Section 4.2, the article proposes fabric storage based on
erasure codes, and stores the training weight parameters and
models of federated learning to Fabric Ledger. We propose to
make the block file BF a fixed size Sm , so that after dividing
m file blocks, each file block is a fixed size Smm byte. The ex-
periment ignores the time when blockfiles are encoded and
decoded through erasure codes, and only considers the en-
ergy consumed by block uploading Fabric. Assuming Sm =1MB,
Sm
m = {8, 16, 32, 64, 128}, Fig. 7 shows the comparison of the
total time of BF ingests on Fabric and retrieve Fabric when
Sm =1MB and BF is divided into five different data block sizes.
From the experimental data, it can be seen that the smaller
a single data block is, the more times it needs to be uploaded
Fig. 8 – The retrieval throughput of Fabric. and retrieved. In this situation, transaction will lead to take
more time for the process of BF upload and verification. With

Fig. 9 – Comparison of training model scores (left → right: RSP-AFI(MLP), RSP-AFI(DT),RandomForest and SVM).
12 computers & security 99 (2020) 102033

Table 4 – The detailed description for KDDCup99 Database.

Attack Category Description Features Train Test

Normal Normal records 97,278 60,593
DoS Denial of service attack f5 f6 f8 f13 f29 f30 f32 f34 f38 f39 391,458 229,853
f3−(ecr−i ) f4−(RSTR ) f4−(S )
0
Probe Monitor and detect other f5 f6 f27 f34 f35 f36 f40 f3−( ft p−dat a ) 4,107 4,166
activities f3−(htt p) f3−( private ) f3−(t el net ) f3−(smt p)
R2L Illegal access to remote f1 f5 f6 f10 f11 f19 f32 f33 f34 f36 f37 1,126 16,189
machines f3−( ft p−dat a ) f3−(other )
U2R Illegal access to local super f1 f5 f6 f10 f14 f17 f18 f24 f32 f34 f37 52 228
user privileges by ordinary f3−( ft p−dat a ) f3−(other )
users
Total 494,021 311,029

the increase of data block capacity, the time of uploading file in this experiment, 13 features are extracted from each attack
information is shorter. By comparing the experimental data, type to show better training results.
we can get the shortest time to trace the BF when the data In the model test score experiment part, we compared
block is 64KB. Therefore, while Smm =64KB=65536 bytes, the to- two supervised learning algorithms and two semi-supervised
tal time required for BF upload and download is the smallest. learning algorithms. This experiment will focus on the train-
The retrieval throughput of Fabric was benchmarking with ing results of the test set in the training model. In the ex-
the Hyperledger Caliper. Fig. 8 shows the throughput changes perimental training, all machine learning algorithms have un-
of retrieval data of different size blocks at different transac- dergone 50 iterations of training, and the training data of the
tion arrival rates. The transaction arrival rate (τ ) is specified semi-supervised learning algorithm is labeled with a 20% la-
in the fabric configuration τ = T/B. T is defined as the num- bel. Fig. 9 is the training scores of KDDCup99 in RSP-AFI (MLP),
ber of transaction. B is defined as limits the minimum time in RSP-AFI (DecisionTree), RandomForest and SVM. Because of
the ordering service for a batch transaction to package a block. the imbalance of the 10% KDDCup99 attack data set, U2R’s
Experimental data show that when the transaction arrival rate alarm traffic is significantly smaller than the other three types
(τ ) is 100 TPS and block size is 128 GB, retrieve the data in the of attacks. Therefore, U2R attacks with a smaller amount of
Fabric of the highest throughput. But still need to consider the data produced higher test scores in different training scores,
actual storage condition, can reduce the retrieval performance so only the other three types of attacks are discussed. In
request, looking for the value of the optimal τ and block size, Fig. 9(a), the MLP algorithm has a test score of 0.893 for Probe
meet the actual needs of the business (Wickboldt, 2019). attack. Among the F1-Score scores in Fig. 9(d), the MLP algo-
The experimental data used KDDCup99 intrusion detec- rithm also produced the highest score of 0.93 for Probe attacks.
tion data set. KDDCup99 is the 9-week network connection It can be seen from other test results that the prediction result
data collected on the simulated US Air Force LAN and is di- of the RSP-AFI module in the MLP algorithm is better than
vided into labeled training data and unlabeled test data. The DecisionTree. The Random Forest algorithm is to train mul-
test data and the training data have different probability dis- tiple decision trees, generate models, and then comprehen-
tributions. The test data contains some types of attacks that sively use multiple decision trees for classification. In most
do not appear in the training data, which makes intrusion de- attack scenarios, the training effect of random forest is better
tection more realistic. The identification types of KDDCup99 than RSP-AFI (DecisionTree). The SVM algorithm is a super-
are shown in Table 4. A network connection of KDDCup99 is vised learning algorithm. It can be learned from the experi-
defined as a sequence of TCP data packets from the beginning mental scores that SVM and RSP-AFI (MLP) have similar train-
to the end within a certain period of time. During this time, ing results. However, the RSP-AFI (MLP) algorithm is a semi-
data is transferred from the source address to the destination supervised learning algorithm, and it is a considerable result
address under a predefined protocol. In this experiment, 10% to obtain a test score similar to the supervised learning algo-
of KDDCup99 network traffic was selected. Among them, the rithm under training with 20% labeled data. In a real applica-
training data set contains 494,021 pieces, among which 97,278 tion scenario, there is no fully marked alarm training set. CIDS
pieces of normal information and 398,743 pieces of intrusion can only mark part of the data based on the alarms that have
data. The test data set is 311029, including 60593 normal data occurred. Therefore, the alarm filtering and identification of
and 250,436 intrusion data. There are 24 attacks in the training RSP-AFI (MLP) is more suitable for real scenarios.
set and 38 attacks in the test set. According to Selvakumar and The ROC (Receiver Operating Characteristic) curve is called
Muneeswaran (2019), KDDCup99 after feature screening can the receiver operating characteristic curve. It was first used
achieve better training scores and improve training speed. So in the field of radar signal detection to distinguish between
signal and noise. ROC is now widely used to evaluate the pre-
 computers & security 99 (2020) 102033 13

based on blockchain ledger.The experimental part evaluates

the scheme from the perspective of the storage efficiency of
the blockchain and the intrusion warning recognition model.
The proposed model was compared with important machine
learning algorithms, and considerable test scores were ob-
tained. The challenges and directions of future research are:
(1) Explore the efficiency of RSP alarm analysis in real scenar-
ios to meet the application requirements of IoT devices; (2)
Set the white/black list mechanism of alarms in CIDS in IoT
devices, and The end device accurately handles the intrusion
crisis and reduces the bandwidth pressure caused by the up-
load of false alarms.

Declaration of Competing Interest

The authors declare that they have no known competing fi-

nancial interests or personal relationships that could have ap-
Fig. 10 – Comparison of AUC between different model peared to influence the work reported in this paper.
algorithms.
CRediT authorship contribution statement

Xinhong Hei: Investigation, Conceptualization, Methodol-

dictive ability of models. AUC (Area Under Curve) is the area of
ogy, Project administration. Xinyue Yin: Software, Validation,
ROC curve. The experiment uses AUC to evaluate the accuracy
Writing - original draft. Yichuan Wang: Data curation, Super-
of the overall model. The experiment tests the overall model
vision, Funding acquisition. Ju Ren: Formal analysis, Visual-
of the scheme from the machine learning algorithms of feder-
ization, Resources. Lei Zhu: Data curation, Writing - review &
ated learning, graph convolutional neural network (Yao et al.,
editing.
2019) and deep neural network (Vinayakumar et al., 2019), and
sets the same learning rate (λ=1). Fig. 10 is a violin chart of
the AUC accuracy distribution of the BFL-CIDS scheme and
Acknowledgments
the other two models. In the 100 tests performed, the aver-
age AUC of the three models are 0.908, 0.849, 0,899. Different
This research work is supposed by the National Key R&D Pro-
types of attacks on the same data set need to extract differ-
gram of China (2018YFB1201500), National Natural Science
ent features, and different accuracy affects the distribution of
Founds of China (61602376, 61773313, 61602374, 61702411),
training results. According to the data distribution in the violin
National Natural Science Founds of Shaanxi (2017JQ6020,
chart, the accuracy of the text-based GCN model is the largest.
2016JQ6041), Key Research and Development Program of
The training results of federated learning and DNN model are
Shaanxi Province (2020GY-039, 2017ZDXMGY-098, 2019TD-
similar and relatively stable. The advantage of federated learn-
014), Science Technology Project of Shaanxi Education Depart-
ing is that it has better privacy protection capabilities. And the
ment (16JK1573, 16JK1552).
semi-supervised learning algorithm is used in the BFL-CIDS
model, which can exert better data processing and prediction
R E F E R E N C E S
capabilities in practical applications.

Abadi M, Chu A, Goodfellow I, McMahan HB, Mironov I, Talwar K,

6. Conclusion Zhang L. Deep learning with differential privacy. In: In
Proceedings of the 2016 ACM SIGSAC Conference on
The BFL-CIDS scheme solves the problem of the data secu- Computer and Communications Security (CCS’16).
rity protection and preventive maintenance of the current Association for Computing Machinery; 2016. p. 308–18.
doi:10.1145/2976749.2978318. New York, NY, USA
cloud intrusion detection system. Several advantages of BFL-
Androulaki E, Manevich Y, Muralidharan S, Murthy C,
CIDS scheme are as follows: (1) Strengthen the privacy pro- Laventman G. Hyperledger fabric: a distributed operating
tection for users: the combination of federated learning and system for permissioned blockchains. the Thirteenth EuroSys
blockchain can fully protect the security of private data; (2) Conference, 2018.
Relieve client pressure: for LN, only a small amount of mem- Benhamouda F, Halevi S, Halevi T. Supporting private data on
ory is occupied, and the calculation and storage pressure are hyperledger fabric with secure multiparty computation. In:
2018 IEEE International Conference on Cloud Engineering
transferred to RSP and GTP; (3) Efficient feedback: accord-
(IC2E); 2018. p. 357–63. doi:10.1109/IC2E.2018.00069. Orlando,
ing to the existing model of permissioned blockchain, RSP
FL
provides timely feedback to some alarms to cope with the Blum A, Mitchell T. Combining labeled and unlabeled data with
complex and changeable network environment. (4) Unified co-training. In: Proc. Workshop Comput. Learn. Theory; 1998.
management: massive alerts samples are stored permanently p. 92–100.
14 computers & security 99 (2020) 102033

Bostani H, Sheikhan M. Hybrid of anomaly-based and privacy-preserving machine learning. In: 2017 IEEE
specification-based IDS for internet of things using Symposium on Security and Privacy (SP), San Jose, CA; 2017.
unsupervised OPF based on mapreduce approach. Comput. p. 19–38. doi:10.1109/SP.2017.12.
Commun. 2017;98:52–71. Rahman MA, Asyhari AT, Leong LS, Satrya GB, Tao MH,
Chai D., Wang L., Chen K., et al. Secure federated matrix Zolkipli MF. Scalable machine learning-based intrusion
factorization[a/OL]. 2019. ArXiv.org(2019-6-12) detection system for IoT-enabled smart cities. Sustainable
http://arxiv.org/abs/1906.05108. Cities andSociety 2020;61:102324.
Cramer R, Damgård I, Escudero D, Scholl P, Xing C. SPDZ2k : doi:10.1016/j.scs.2020.102324. ISSN 2210-6707
Efficient MPC mod 2k for dishonest majority 2018;2018:482. Ramanan P., Nakayama K.. Baffle : blockchain based aggregator
doi:10.1007/978-3-319-96881-0_26. free federated learning. 2019. ArXiv.org(2019)
Geyer R.C., Klein T., Nabi M.. Differentially private federated https://arxiv.org/abs/1909.07452.
learning: a client level perspective. 2017. ArXiv.org(2017) Ren J, Guo H, Xu C, Zhang Y. Serving at the edge: A scalable IoT
https://arxiv.org/abs/1712.07557. architecture based on transparent computing. IEEE Network
Gupta H, Hans S, Mehta S, Jayachandran P. On building efficient 2017;31(5):96–105. doi:10.1109/MNET.2017.1700030.
temporal indexes on hyperledger fabric. In: 2018 IEEE 11th Ren J, Zhang Y, Zhang K, Shen X. Adaptive and channel-aware
International Conference on Cloud Computing (CLOUD), San detection of selective forwarding attacks in wireless sensor
Francisco, CA; 2018. p. 294–301. networks. IEEE Transactions on Wireless Communications
Hei X, Bai B, Wang Y, Zhang L, Zhu L, Ji W. Feature extraction 2016;15(5):3718–31. doi:10.1109/TWC.2016.2526601. May
optimization for bitstream communication protocol format Selvakumar B, Muneeswaran K. Firefly algorithm based feature
reverse analysis. In: 2019 18th IEEE International Conference selection for network intrusion detection. Computers &
On Trust, Security And Privacy In Computing And Security 2019;81:148–55. doi:10.1016/j.cose.2018.11.005. ISSN
Communications/13th IEEE International Conference On Big 0167-4048
Data Science And Engineering (TrustCom/BigDataSE), Rotorua, Sharma PK, Park JH, Cho K. Blockchain and federated
New Zealand; 2019. p. 662–9. learning-based distributed computing defence framework for
doi:10.1109/TrustCom/BigDataSE.2019.00094. sustainable society. Sustainable Cities andSociety
Li W, Meng W, Au MH. Enhancing collaborative intrusion 2020;59:102220. doi:10.1016/j.scs.2020.102220. ISSN 2210-6707
detection via disagreement-based semi-supervised learning Soldatos J, Kefalakis N, Hauswirth M, Serrano M, Herzog R.
in IoT environments. Journal of Network and Computer Openiot: open source internet-of-things in the cloud. Lecture
Applications 2020;161:102631. Notes in Computer Science 2015;9001:13–25.
Lindell Y. In: Encyclopedia of Data Warehousing and Mining. Thai L., Jonathan S., Steven B., Code sample: Intel ISA-L erasure
Secure multiparty computation for privacy preserving data code and recovery[E B/OL]. 2015. (2015-4-16)
mining; 2006. doi:10.4018/9781591405573.ch189. https://software.intel.com/zh-cn/articles/
Lu Y, Huang X, Dai Y, Maharjan S, Zhang Y. Blockchain and intel- isa- l- erasure- code- and- recovery.
federated learning for privacy-preserved data sharing in Tian Y, Wang Z, Xiong J, Ma J. A blockchain-based secure key
industrial iot. IEEE Transactions on Industrial Informatics management scheme with trustworthiness in DWSNs. IEEE
2019;PP(99). 1-1 Transactions on Industrial Informatics 2020;16(9):6193–202.
Lu Y, Huang X, Dai Y, Maharjan S, Zhang Y. Differentially private doi:10.1109/TII.2020.2965975. Sept.
asynchronous federated learning for mobile edge computing Viinikka J, Debar H, Ludovic, et al. Time series modeling for IDS
in urban informatics. IEEE Transactions on Industrial alert management[c]. In: ACM Symposium on Information
Informatics 2020;16(3):2134–43. doi:10.1109/TII.2019.2942179. Computer and Communications Security. ACM; 2006.
March p. 102–13.
Ma C., Li J., Ding M., Yang H.H., Shu F., Quek T.Q.S., et al. On Vinayakumar R, Alazab M, Soman KP, Poornachandran P,
safeguarding privacy and security in the framework of Al-Nemrat A, Venkatraman S. Deep learning approach for
federated learning. 2019. ArXiv.org(2019) intelligent intrusion detection system. IEEE Access
https://arxiv.org/abs/1909.06512v2. 2019;7:41525–50. doi:10.1109/ACCESS.2019.2895334.
Majeed U, Hong CS. FLchain: Federated learning via MEC-enabled Wang S.. BlockfedML: Blockchained federated machine learning
blockchain network. In: 2019 20th Asia-Pacific Network systems. 2019. 751–756, 10.1109/ICICAS48597.2019.00162
Operations and Management Symposium (APNOMS), Matsue, Wang Y, Bai B, Hei X, Zhu L, Ji W. An unknown protocol syntax
Japan; 2019. p. 1–4. doi:10.23919/APNOMS.2019.8892848. analysis method based on convolutional neural network.
Martinez I, Francis S, Hafid AS. In: CyberC 2019 Workshop on Trans. Emerging Tel. Technol. 2020:e3922. doi:10.1002/ett.3922.
Blockchain. Record and reward federated learning Wang Y, Zhu H, Hei X, Kong Y, Ji W, Zhu L. An energy saving based
contributions with blockchain; 2019. on task migration for mobile edge computing. EURASIP
doi:10.1109/CyberC.2019.00018. Journal on Wireless Communications and Networking
Meidan Y, Bohadana M, Mathov Y, Mirsky Y, Breitenbacher D, 2019;2019. doi:10.1186/s13638-019-1469-2.
Shabtai A, et al. N-baiot: network-based detection of iot Wickboldt C.. Benchmarking a blockchain-based certification
botnet attacks using deep autoencoders. IEEE Pervasive storage system. 2019. 10.13140/RG.2.2.32684.31360
Computing 2018;17(3):12–22. JulSep Xie P., Bilenko M., Finley T., Gilad-Bachrach R., Lauter K., Naehrig
Meng Y, Kwok L. Practical Applications of Intelligent Systems M.. Crypto-nets: Neural networks over encrypted data. 2014.
Advances in Intelligent and Soft Computing, vol 124. In: ArXiv preprint arXiv preprint arXiv:1412.6181.
Wang Y, Li T, editors. Adaptive false alarm filter using Yao L, Mao C, Luo Y. Graph convolutional networks for text
machine learning in intrusion detection. Berlin, Heidelberg: classification, 33; 2019. p. 7370–7.
Springer; 2011. Yin B, Yin H, Wu Y, Jiang Z. FDC: A secure federated deep learning
Midi D, Rullo A, Mudgerikar A, Bertino E. Kalis - a system for mechanism for data collaborations in the internet of things.
knowledge-driven adaptable intrusion detection for the IEEE Internet ofThings Journal 2020.
internet of things. In: 2017 IEEE 37th International Conference doi:10.1109/JIOT.2020.2966778.
on Distributed Computing Systems (ICDCS); 2017. p. 656–66. Zarpelo BB, Miani RS, Kawakani CT, de Alvarenga SC. A survey of
June intrusion detection in internet of things. Journal of Network
Mohassel P, Zhang Y. SecureML: a system for scalable and Computer Applications 2017;84:25–37.
 computers & security 99 (2020) 102033 15

Zhang Y, Lu Y, Huang X, Zhang K, Maharjan S. Blockchain Ju Ren received the B.Sc. (2009), M.Sc. (2012), Ph.D. (2016) degrees
empowered asynchronous federated learning for secure data all in computer science, from Central South University, China.
sharing in internet of vehicles. IEEE Transactions on Vehicular During 2013-2015, he was a visiting Ph.D. student in the Depart-
Technology 2020;PP(99). 1–1 ment of Electrical and Computer Engineering, University of Water-
Zhou ZH, Li M. Tri-training: exploiting unlabeled data using three loo, Canada. Currently, he is a professor with the School of Com-
classifiers. IEEE Transactions on Knowledge & Data puter Sci-ence and Engineering, Central South University, China.
Engineering 2005;17(11):1529–41. His research in-terests include Internet-of-Things, wireless net-
Zhu H, Wang Y, Hei X, Ji W, Zhang L. A blockchain-based working systems, net-work computing and edge computing. Dr.
decentralized cloud resource scheduling architecture. In: 2018 Ju Ren has published more than 70 papers in top journals and
International Conference on Networking and Network conferences, including IEEE JSAC, TIFS, TMC, TCC and IEEE IN-
Applications (NaNA), Xi’an, China; 2018. p. 324–9. FOCOM, ICDCS, etc. He is a recipient of the best paper award of
doi:10.1109/NANA.2018.8648712. IEEE ICC’19 and IEEE IoP’18, the outstanding paper award of IEEE
HPCC’19 and the most popular paper award of Chinese Journal of
Xin-Hong Hei received his B.S. degree and M.S. de-gree in com- Electronics (2015-2018). He currently serves/has served as an as-
puter science and technology from Xi’an University of Technol- sociate editor for IEEE Transactions on Vehicular Technology and
ogy, Xi’an, China, in 1998 and 2003, respectively, and his Ph.D. de- Peer-to-Peer Networking and Applications, a guest editor for IEEE
gree from Nihon University, Tokyo, Japan, in 2008. He is currently a Transactions on Industrial Informatics and IEEE Network, and a
professor with the Faculty of Computer Science and Engineering, TPC member of many international con-ferences including IEEE
Xi’an University of Technology, Xi’an, China. His current research INFOCOM’20/19/18, Globecom’17, WCNC’17, WCSP’16, etc. He also
in-terests include intelligent systems, safety-critical system, and served as the TPC chair of IEEE BigDataSE’19, a poster co-chair of
train con-trol system. IEEE MASS’18, a track co-chair for IEEE/CIC ICCC’19, IEEE I-SPAN’18
and VTC’17 Fall, and an active reviewer for over 20 international
Xin-Yue Yin is studying for master’s degree in computer applica-
journals. He is a member of IEEE and ACM.
tion in Xi’an University of technology. Her research area is network
security and blockchain. Lei Zhu received his Ph.D. degree in computer science and tech-
nology from Xi’an Jiaotong University, China, in 2014. He is cur-
Yi-Chuan Wang received his Ph.D. degrees in computer system ar-
rently working as Xi’an University of Technology, Department of
chitecture from Xidian University of China in 2014. He is an ACM
computer science and engineering. His research interests include
member and a CCF member. Now he is a Lecturer in Xi’an Uni-
secret sharing scheme, data mining and graph mining.
versity of Technology and with Shaanxi Key Laboratory of Network
Computing and Security Technology. His research areas include
cloud computing and networks security.