You are on page 1of 11

IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO.

1, JANUARY 1, 2021 585

False Data Injection Attack Detection for Industrial


Control Systems Based on Both Time- and
Frequency-Domain Analysis of Sensor Data
Dajian Huang, Xiufang Shi , Member, IEEE, and Wen-An Zhang , Member, IEEE

Abstract—This article studies the intrusion detection problem remote sensing sensors used in remote control systems are a
for industrial control systems (ICSs) with repetitive machining good example [2]. In the context of the industrial Internet, IoT
under false data injection (FDI) attacks. A data-driven intrusion technology has promoted the development of ICS toward a dis-
detection method is proposed based on both time- and frequency-
domain analysis. The proposed method only utilizes the sensor tributed and edge-cloud collaboration architecture. Meanwhile,
measurements required in closed-loop control, and does not con- ICS is also an application of IoT and the original closedness
sume additional system resources or rely on the system model. In of ICS is broken. Various viruses, e.g., Trojans intrude the ICS
addition, features in time and frequency domain are extracted at within normal information flow, which seriously threatens the
the same time, having higher reliability than the intrusion detec- security of ICS. In recent years, ICS security incidents have
tion methods which only utilize the features in time domain.
After feature extraction, hidden Markov models (HMMs) are also increased, thus it is of great importance to design secu-
established by using the feature vectors under normal operating rity protection schemes for ICS [3]–[5]. False data injection
conditions of the ICS, and then the trained HMMs are utilized in (FDI) attacks [6] can destroy the normal processing of the
real-time intrusion detection. Finally, experiments are carried out ICS by tampering with the data transmitted in the network.
on a networked multiaxis engraving machine with FDI attacks. It can be implemented by man-in-the-middle (MITM) attacks
The experimental results show the effectiveness and superiority
of the proposed intrusion detection method. and internal attacks, and has good concealment. For exam-
ple, the famous “Stuxnet virus” was implemented by internal
Index Terms—False data injection (FDI) attacks, hidden attacks [7]. FDI attacks are one of the main security threats
Markov model (HMM), industrial control systems (ICSs), intru-
sion detection, signal analysis. of current ICS.
ICS security protection has received widespread attentions
from both industry and academia. People began to learn
I. I NTRODUCTION from the field of traditional information security to solve
ICS security protection issues, such as industrial firewalls,
HE EARLY industrial control systems (ICSs) were rela-
T tively independent and isolated from the external Internet.
However, with the rapid development and applications of the
ICS vulnerability mining, and risk assessment [8]–[10]. While
vulnerability mining and risk assessment cannot provide real-
time network security protection. Advanced attackers can find
industrial communication networks, ICS is gradually develop- the vulnerabilities in the firewall, and bypass the firewall for
ing toward a networked and open architecture. The centralized attacks. Besides, the firewall cannot defend against internal
control structure has been transformed into the current dis- attacks. Intrusion detection is promising to ensure system
tributed control structure. Many computation and information security through system monitoring and abnormal warning.
transmission functions have been gradually decentralized and It collects information about devices and networks in ICS.
spread from the central node to the edge nodes, that is, edge The data are analyzed and identified to determine whether
computing (EC), where resource provisioning is pushed from there is abnormal behavior in the system. Therefore, as an
the cloud infrastructure to the edge of the Internet [1]. This is active protection technology for system security, the intrusion
right because of the development of Internet-of-Things (IoT) detection method can effectively make up for the shortcomings
technology that each node in ICS, such as sensor, controller, or of traditional security protection technologies, such as a fire-
actuator, can be equipped with IoT modules, thus having the wall. It is regarded as the second security defense line of ICS,
ability of computation and information transmission. Industrial which can realize real-time detection of external and internal
intrusions [11].
Manuscript received April 23, 2020; revised May 24, 2020; accepted
June 22, 2020. Date of publication July 7, 2020; date of current ver- Traditional information technologies are used to solve the
sion December 21, 2020. This work was supported in part by the Key intrusion detection issues of ICS, including network-based
Research and Development Program Projects in Zhejiang Province under and host-based (device-based) methods. The network-based
Award 2019C03098, and in part by the National Natural Science Foundation
of China under Grant 61822311 and Grant 61801422. (Corresponding author: intrusion detection obtains the features of the key field by
Wen-An Zhang.) acquiring the information in the network information flow,
The authors are with the College of Information Engineering, and performs real-time detection using data packets, traffic,
Zhejiang University of Technology, Hangzhou 310023, China (e-mail:
humdrumj@163.com; xiufangshi@zjut.edu.cn; wazhang@zjut.edu.cn). protocols, and topology structures in the network [12]–[16].
Digital Object Identifier 10.1109/JIOT.2020.3007155 For FDI attacks, there is no significant change in network
2327-4662 
c 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
586 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 1, JANUARY 1, 2021

traffic. Besides, the operational behaviors of the monitored


ICS are not fully accessible to the network-level information.
In some cases, if the cyberattack behavior does not violate
the protocol used by ICS or the communication mode among
devices, the network-based intrusion detection system will not
be able to detect such attacks. The host-based intrusion detec-
tion methods check the audit of the host device, the process
status of the device, the log content, and attributes, such as
the login status of the device, the operation authority of the
operator on the device and the file, the sensitive operation
performed by the operator, the operation variables and out-
put variables of the device, operating states of device and
other information, based on which, the anomalous behaviors Fig. 1. Typical structure of networked ICS with repetitive processing.
are identified [17]–[26]. However, this kind of methods con-
sumes a lot of system resources in acquiring and analyzing
the above-mentioned data, and may not meet the real-time is feasible to implement intrusion detection by analyzing the
requirements of the ICS due to weak computing and storage system output from sensor measurements.
capabilities of the equipment [27]. Moreover, due to the differ- To this end, this article proposes an ICS intrusion detec-
ences in the structure and functions between various ICSs, the tion method based on the analysis of sensor measurements.
host-based intrusion detection methods have low portability. The proposed method includes three parts: 1) zero-mean pro-
Although many kinds of traditional information technologies cessing of the sensor data to obtain the deviation between the
have been designed in the regular IT system, these technolo- actual trajectory and the expected trajectory; 2) feature extrac-
gies cannot be applied in ICSs without any change due to tion from both the time and frequency domain; and 3) hidden
the industrial-oriented communication protocols or the unique Markov models (HMMs) training using the extracted features
properties [26]. under normal operating conditions and real-time intrusion
For the ICS, if the devices, e.g., sensors or controllers are detection using the trained HMMs. The main contributions
attacked or the network transmission is destroyed or falsi- of this article are summarized as follows.
fied, the input and output signals of the controlled object 1) Consider the ICS with repetitive processing, a data-
will eventually deviate from the normal working condition. driven intrusion detection method using only sensor
Xu et al. [28] has experimentally verified that the attacks measurements is proposed. This method does not rely
may cause blindness and malfunction of sensors. From this on the system model. Moreover, since the utilized
point of view, the result is similar to a system fault (such as sensor measurements are the system output in closed-
a sensor fault). Therefore, many researchers use fault diag- loop control, there is no additional system resource
nosis and bad data detection to solve the intrusion detection consumption.
issues of ICS, such as binary hypothesis detector under the 2) For the sensor measurements, features in both time and
Bayesian framework [29], [30], least-squares detector [31], frequency domain are extracted at the same time, they
χ 2 -detector based on the statistical feature analysis [32], are more reliable than the time-domain features utilized
and residual generation principle-based detector [33]–[35]. in the existing intrusion detection methods.
However, most of these detectors need a precise ICS model, 3) Experiments are carried out on a self-developed net-
whose accuracy will affect the accuracy and reliability of worked multiaxis engraving platform. The experimental
intrusion detection. In addition, once the attacker knows results show the effectiveness and superiority of the
the model parameters of the ICS, it can design a covert proposed intrusion detection method.
FDI attack and successfully escape these intrusion detection The remainder of this article is organized as follows. Section II
methods [6]. describes the FDI intrusion detection problem to be solved.
Therefore, how to design an intrusion detection method that Section III introduces the proposed HMM-based intrusion
neither depends on the ICS model nor needs additional system detection method. Section IV shows the experimental results.
resources is particularly important. Note that in the closed- Section V concludes this article and provides some research
loop feedback control of ICS, the system needs to acquire directions in the future. Notations in this paper are presented
the system output measured by various sensors. If we can in Table I.
directly utilize these sensor measurements in intrusion detec-
tion, there would be no requirement on system resources or II. P ROBLEM S TATEMENT
system models. Besides, the signals transmitted through the We consider a networked ICS with repetitive processing
network in the ICS include the reference signals sent from the as shown in Fig. 1, which is very common in industrial
server to the controller, the control signals sent from the con- systems. The server sends the reference trajectory to the field
trollers to the actuators, and the output signals sent from the controller through the network, the controller (proportion inte-
sensors to the controllers. No matter what data are modified gration differentiation) calculates the control input according
by the attackers, it will eventually affect the system output to the deviation between the reference trajectory and the actual
because the system is closed-loop controlled. Therefore, it machining trajectory measured by the sensors, then sends the
HUANG et al.: FALSE DATA INJECTION ATTACK DETECTION FOR INDUSTRIAL CONTROL SYSTEMS 587

TABLE I
N OTATIONS other factors, and the interference is an attack signal added
by the attacker, such as a sine signal. To ensure concealment,
adversaries may set the amplitude of η(t) to be much smaller
than the amplitude of g(t).
The objective of the article is to design a data-driven FDI
attack detection method. The proposed method implements the
detection of FDI attacks only by using the measurements f (t)
of the actual machining trajectory, no explicit mathematical
model of the controlled object is required. We will extract
the time- and frequency-domain features of the sensor mea-
surements, and then design an appropriate machine learning
method to achieve real-time intrusion detection. The designed
method needs to meet the safety and real-time performance
requirements of the ICS.

III. P ROPOSED I NTRUSION D ETECTION M ETHOD


The idea of the proposed intrusion detection method is
shown in Fig. 2. The data f (t) are processed by the zero-
mean method first; and then the signal processing technique
is employed to extract the features in both time and frequency
domain; finally, the computationally efficient HMM is used
for intrusion detection.

control input to the actuator through the field bus to execute the A. Data Processing
control input. During the data transmission between the server When the system is under attack, the measured data will
and the controller, it may be subject to MITM attacks from be changed due to the injected data. Since the amplitude of
the network or internal attacks by internal operators. These η(t) is much smaller than the amplitude of g(t), g(t) is similar
internal attacks and MITM attacks can inject false data into to f (t). In such condition, directly distinguishing between the
the reference trajectory and control input. Whether the refer- measured data stream f (t) and g(t) is difficult. It is necessary
ence trajectory or the control input is modified, it will make the to extract the features effectively from the data for intrusion
actual machining trajectory deviate from the normal value. In detection.
this article, our purpose is to implement intrusion detection by In order to eliminate the impact of reference trajectory
analyzing the data of the actual machining trajectory measured signal s(t) of intrusion detection, s(t) is separated from the
by sensors. measurement signal f (t), and the residual information ξ(t) or
The reference trajectory at the t-sampling moment is ξ(t) + η(t) in the normal state is compared with that under
denoted by s(t). Due to the measurement errors of the sen- attack. The zero-mean method is employed to implement this
sors, the actual measured machining trajectory values at the idea, as is shown in Fig. 3, where the data at the 800th point
t-sampling time can be represented as of 100 periods of the experimental data are presented. It is
noted that the reference trajectory is periodic, and the actual
g(t) = s(t) + ξ (t) (1) measurements consist of reference trajectory and noise. The
where ξ(t) is the measurement noise, which is generally data at the same sampling points in different periods are shown
assumed to follow a zero-mean white Gaussian distribution in the bottom left of Fig. 3, which are approximated as noise
N(0, σ 2 ), and σ 2 is the noise variance. with a certain mean value, and is conformed to the normal
In this work, adversaries may launch attacks by tamper- distribution law as shown in the bottom right of Fig. 3.
ing with sensor measurements. The output of the closed-loop To preserve the attack signal and eliminate other signals
control will then be abnormal due to the tampered measure- in the data, the following operations are performed. First, the
ments. If the system is subject to FDI attacks, the sensor mean values of the data over multiple periods are calculated
measurements will be as follows:
N
f (t) = s(t) + ξ (t) + η(t) (2) gn (t)
f̄ (t) = n=1 (3)
 N 
where η(t) represents the interference added to the sen- F̄ = f̄ (1), f̄ (2), . . . , f̄ (T) (4)
sor measurements by the attacker, f (t) represents tampered
measurements. Since the considered system has a repetitive where T is the period length, t = 1, . . . , T, N is the number
process, the reference trajectory repetition period is set as T, of periods, gn (t) is the measurements at the t sample points in
namely, s(t) = s(t + T). the nth period, f̄ (t) is the corresponding mean value of g(t) at
It is worth noting that the noise is caused by environmental the t sample points, f̄ (t) = f̄ (t + T), and F̄ is the mean vector
factors, mechanical resonance, sensor measurement errors or of the measurements in N periods.
588 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 1, JANUARY 1, 2021

Fig. 2. Intrusion detection using time- and frequency-domain analysis.

(a)

(b)

Fig. 4. Schematic of feature extraction using the sliding window technique.


Fig. 3. Gaussian properties of periodic data.
(a) Feature extraction in time domain. (b) Feature extraction in frequency
domain.

Second, the mean values are subtracted from the new data, by a feature vector Etime (t) in time domain, and is quantized
namely, zero mean, to obtain the error signal as follows: manually according to the range of actual feature vector (for
example, a value between 0.000 and 0.001 is quantized to 1,
R(t) = f (t) − f̄ (t). (5) and a value between 0.001 and 0.002 is quantized to 2, etc.).
By subtracting the mean f̄ (t) from the new data, the data we The quantized values are taken as the input of HMM, and
obtain are zero mean. With this step, the original data stream the number of species of the quantized values is the obser-
is transformed into the error signal, which will eliminate the vation number. The sliding window technique is utilized in
influence of the original signal on intrusion detection. HMM training and real-time intrusion detection. The length is
chosen as L1 and step size is set as 1. The schematic about
feature extraction using the sliding window technique in the
B. Feature Extraction time domain is shown in Fig. 4(a).
After zero-mean data processing, the proportion of the 2) Frequency-Domain Analysis: As mentioned above, the
attack signal in the error signal is larger than that in the mea- error signal after zero-mean processing should be analyzed
sured data, and it is easier to extract the features in the time not only in time domain but also in frequency domain. In the
domain to find out the abnormal situation. In the time domain, frequency-domain analysis, a fast discrete Fourier transform
however, the features are limited, and some attacks may escape (FFT) combined with the sliding window technique is utilized.
from time-domain analysis (for example, a sinusoidal attack The length of the sliding window is chosen as L2 and the
whose amplitude is almost the same as the error signal, but step size is set as 1. For the error signal R(t) containing L2
with a very high frequency, which will affect the actuators uniformly sampled values f (t), the FFT is as follows:
and controllers after a long run). In order to extract more
2 −1
L
comprehensive features, the information in the time domain − 2π jk
L i
Y(k + 1) = e 2 R(j + 1) (6)
is combined with that in the frequency domain.
j=0
1) Time-Domain Analysis: Discrete HMM is considered for
training and intrusion detection. The error signal is represented where j, k = 0, 1, . . . , L2 − 1.
HUANG et al.: FALSE DATA INJECTION ATTACK DETECTION FOR INDUSTRIAL CONTROL SYSTEMS 589

The mean is calculated for the vertical axis value cor-


responding to the integer k, and it is represented as
the frequency-domain feature vector Efrequency (t) at current
moment t
L2 −1
|Y(k + 1)|
Efrequency = h=0 . (7)
L2
(a)
The reason why Efrequency (t) is chosen as the feature vector is
that no matter what the attack frequency is, the feature vector
will change as long as the attacks are added into the system.
After the operation of sliding window (wherein the length of
window is L2 , the step size is 1), the feature in the frequency
domain corresponding to the time will be obtained. Then the
fast and efficient HMM is employed to intrusion detection.
The diagram of the feature extraction using sliding window
technique in the frequency domain is shown in Fig. 4(b). (b)

C. Intrusion Detection Fig. 5. Schematic of model training and output. (a) Training for HMM
parameters. (b) Evaluating possibilities under model parameters.
1) Model Training: HMM [36] is a probability model and
a time-series model, which predicts or classifies states accord-
ing to events that happened in a previous period of time, and in Fig. 5(b). It can be calculated by forward/backward algo-
can obtain better results than just using current information. rithms. Under normal conditions, the probability assessments
Since the forms of potential attacks are unknown, it is diffi- are generally stable and large. If the system is under attack,
cult for us to obtain all kinds of samples under attacks, and the distribution of the observations will be different. In this
the data become imbalanced seriously. HMM can avoid the case, the probability assessments are reduced, i.e., the likeli-
problem of data imbalance because it only models the data in hood that these observations conform to this HMM is reduced.
the normal state and has no need to label the data. It outputs For multiple HMM detectors corresponding to multiple sen-
the probability value according to the relationship between sors, the output of any HMM shows that the data have been
the real-time input sequence and the trained model. Finally, tampered and indicates that the system is under attack.
we compare the probability value with the threshold to judge 2) Design of μ-TF-HMM Detector: The μ-TF-HMM intru-
whether the system is under attack. It is noted that HMM is fast sion detector for an ICS with a single sensor is shown in
and efficient, therefore it is very suitable for the application Fig. 6. Detectors are established for each sensor, for identi-
of ICS. fying the attacked sensor. The abnormality detected by any
An HMM usually consists of the following five-tuple: μ-TF-HMM detector indicates that the ICS is under attack.
The model training and real-time detection are summarized,
λ = {V, M, A, B, C} (8)
respectively, in Algorithms 1 and 2.
where V represents the number of hidden states, M repre- In Algorithm 1, step 1 is to initialize the model parameters,
sents the number of observable states, A = {aij } is the transfer steps 2–6 are to extract the features efficiently, the calcu-
matrix of hidden states, representing the probability of trans- lated mean F̄ in step 2 needs to be stored for the next step
ferring from state i to state j, B = {bi (o)} is a matrix of and Algorithm 2, steps 7 and 8 are to acquire the trained
observable probabilities (confusion matrix), representing the model parameters and threshold, the trained model parame-
probability that observations o are observed from a hidden ters and threshold should also be stored for Algorithm 2. In
state i, and C = {Ci } is the initial state probability, repre- Algorithm 2, steps 1–4 are to extract the features efficiently,
senting the probability vector starting from each hidden state. the L1 in step 5 should be the same to L1 in Algorithm 1,
Both the transition matrix of hidden states and the confusion and the rest steps are to evaluate the probability and report
probability matrix are time independent, that is, these matrices the detection result.
do not change with time. For an HMM with fixed V and M,
λ = {A, B, C} can also be used to represent the parameters of
IV. E XPERIMENT
HMM.
In the proposed method in this article, HMM plays a role in A. Platform
evaluating whether the sensor data are abnormally tampered. The experimental platform is a self-developed networked
The HMM is obtained by training the data under normal oper- multiaxis engraving machine with a network attack kit, as
ating conditions using the Baum–Welch algorithm, where the shown in Fig. 7. The intrusion detection system is com-
model parameter λ represents the normal state, as shown in posed of a host computer [human–machine interface (HMI)],
Fig. 5(a). The HMM probability assessment P(O|λ) repre- an ARM single-chip microcomputer, a servo system, a CAN
sents the probability of the input observations under the model bus, Ethernet, an attack kit, and an intrusion detector. The
parameters, and it will be the evaluation results, as shown main functions of the host computer are receiving data from
590 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 1, JANUARY 1, 2021

Fig. 6. Intrusion detection flow for an ICS with single sensor.

the ARM microcomputer through the TCP/IP protocol, per- inputs or sensor measurements. In this experiment, we con-
forming the control algorithm and sending the control input sider the case that the sensor measurements will be modified.
to the ARM microcomputer. The ARM microcontroller acts The attack is covert and not easy to be seen visually through
as a data transfer station in the system, receives the control the HMI monitoring system, this can be realized by, for exam-
input sent from the host computer, and then sends the con- ple, adding sinusoidal signals with small magnitude into the
trol input to each servo system via the CAN bus, meanwhile, measurements. However, such covert attacks may destroy the
it also acquires the speed, position, and torque of each servo actuator in the control systems significantly.
system on the CAN bus. This key information is encapsulated
and sent back to the host computer for processing. The main B. Data Resource
work of the CAN bus is to realize communication between To verify the performance of the proposed method, data are
the ARM microcontroller and the servo system, and to trans- sampled from the networked multiaxis synchronous engraving
fer important data in real time. The main work of the servo machine in real time for experiments. The attack is performed
system is to respond to the control input from the host com- by adding a sinusoidal signal into the sensor measurements.
puter in real time and make corresponding operations. The In this experiment, one sensor is taken as an example. As
HMI includes visualized position information of the engraving long as the value of each sensor is detected, the attack can be
tool, visualized control input (speed information) of each axis, located. The data are divided into two parts. In the first part,
visualized output values (position information) of each axis, all the data are acquired under normal conditions. The sine
and some operation buttons. The attack kit implements attacks signal is taken as an example, other periodic signals are also
by means of MITM attacks, which can modify the control applicable. The amplitude of the sinusoidal signal is 10 cm and
HUANG et al.: FALSE DATA INJECTION ATTACK DETECTION FOR INDUSTRIAL CONTROL SYSTEMS 591

Algorithm 1 HMM Training Based on Features in Time and


Frequency Domain
Input: Normal sensor data N ∗ T, N periods, T data in each
period
Output: Model parameters, threshold, mean vector F̄
1: Initialize the model parameters
2: Calculate the mean F̄ according to equations (3)-(4)
3: Process the data with Zero-mean according to equation
(5)
4: Use sliding window technique on the data of Step3,
perform Fourier transform on the data in the window
5: Extract the features of the frequency-domain according to
equation (7)
6: Quantify the data obtained by Step3 and Step5 as input
values Otime , Ofrequency for the model
Fig. 7. Networked multiaxis synchronous engraving machine experimental
7: Train for model parameters λtime , λfrequency with Baum-
platform with attack kit.
Welch algorithm
8: Use sliding window technique,  and
 calculate
P(Otime |λtime ) and P Ofrequency λfrequency separately.
According to the practical experience, the threshold 1) Attack 1: Sinusoidal signal with a frequency of 10 Hz
  can and amplitude of 0.01 cm.
be taken as Th_time  = max
  −log(P(O
 time |λ 
time )) and
Th_frequency = max −log P Ofrequency λfrequency . 2) Attack 2: Sinusoidal signal with a frequency of 10 Hz
and amplitude of 0.1 cm.
3) Attack 3: Sinusoidal signal with ac frequency of 0.2 Hz
and amplitude of 0.1 cm.
Algorithm 2 Real-Time Intrusion Detection The mean vector F̄ is calculated by using 240 periodic
Input: Real-time sensor data, mean vector F̄, model param- data. The number of hidden states in HMM is set as V = 3,
eters, threshold and the number of observable states is set as M = 13. The
Output: Detected result Baum–Welch algorithm is used to train the model, where the
FOR each new sample value: maximum number of iteration is set as 1000 and the minimum
1: Calculate error signal R according to formula (5) iteration error is set as 1e–6. The length of the observation
2: Perform Fourier transform on the previous L2 data of sequence is L1 = 30. The length of the Fourier transform
Step1 window is L2 = 128. The detection results are shown in Fig. 9.
3: Extract the features of the frequency-domain according to It can be seen from Fig. 9(a) that all the points under
formula (7) Attack 1 are not successfully detected, only a few points
4: Quantify the data of Step1 and Step3 as input values of under Attacks 2 and 3 are detected, which means the detec-
the model tion performance is very poor. In Fig. 9(b), the detection
5: Take the previous L1 observations to calculate the  evalua- performance is better, while there are still many missed
tion probability P(Otime |λtime ) and P Ofrequency λfrequency detections and a few false detections. In Fig. 9(c), the detec-
respectively tion performance is also better than that in Fig. 9(a). There
6: Compare the evaluation probability with threshold Th_time are fewer missed detections than that in Fig. 9(b), while
and Tfrequency  at the same time, there are a few more false detections.
 
7: if P(Otime |λtime ) > Th_time or P Ofrequency λfrequency > Overall, the performance of μ-TF-HMM and μ-HMM are bet-
Th_frequency then ter than HMM. The proposed method μ-TF-HMM seems more
8: result = 1 (Under attack) sensitive than μ-HMM.
9: else
10: result = 0 (Normal)
C. Effect of Window
To compare the results more specifically, the length of slid-
ing window L1 is changed while other parameters are not.
the period is 8 s. The data sampling interval is 5 ms. It should Accuracy, sensitivity, and F1 score are taken as the evaluation
be noted that natural disturbance exists in the acquired data, criteria, which are, respectively, defined as follows.
and the data are not denoising to avoid removing attacks. This 1) Accuracy = (TP + TN)/(P + N), where TP is true
part is used to train the detector. In the second part, the data positive, TN is true negative, P is positive, and N is
contain the attacked signals, and are used to test the detection negative. Generally, the higher the accuracy, the better
performance. As shown in Fig. 8, data in the first 12 periods the classifier.
are normal, data in the last 30 periods are under attacks. The 2) Sensitivity = TP/P, it represents the proportion of all the
system is subject to the following three sinusoidal attacks. positive examples being classified correctly, and weighs
592 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 1, JANUARY 1, 2021

Fig. 8. Experimental data. Normal: Sampling data without attacks, Attack 1: Sinusoidal signal with a frequency of 10 Hz and amplitude of 0.01 cm, Attack 2:
Sinusoidal signal with a frequency of 10 Hz and amplitude of 0.1 cm, Attack 3: Sinusoidal signal with a frequency of 0.2 Hz and amplitude of 0.1 cm. The
figures in the upper rectangular box are the local zoom result.

Fig. 9. Comparison diagram of detected results. In the figure, the red line shows the moment that attack occurs, the blue line shows the detection results,
where 1 means the system is under attack, and 0 means there is no attack. (a) “HMM” [37] represents an HMM detector where the data are only quantified
(b) “μ-HMM” (our method 1) represents an HMM detector where the data are processed with zero mean. The proposed method (c) “μ-TF-HMM” in this
article (our method 2) represents an HMM detector, where the data are processed with zero mean and combined with signal analysis.

the classifier’s ability to recognize positive examples. work. Moreover, the length of the window will affect the detec-
Here, attacked data are considered as positive examples. tion performance, and both the detection performance and the
3) F1 score = 2TP/(2TP + FP + FN), where FP is false computational capability should be considered when choosing
positive and FN is false negative. F1 score is the har- the length.
monic value of the recall rate and the accurate rate.
When both the recall rate and accurate rate are high, D. Effect of Frequency and Amplitude
the F1 score will be high. To test the detection performance of the proposed detector
The comparison results are shown in Fig. 10. The detection under attacks with different frequencies (ω) and amplitudes
performance of μ-HMM is significantly improved after zero (A), multiple frequencies and amplitudes are tested. Validation
mean. The detection performance can be further improved by is performed with 83 200 measured data, 41 600 of which are
combining features in the frequency domain or increasing the added by the sinusoidal attacks. The corresponding parameters
length of the observation sequence, and it almost converges are as follows.
after the accuracy reaching a certain level. It is worth noting 1) L1 = 30, the amplitude of the attack signal is 0.01 cm,
that, the application of the sliding window technique using the frequency of the attack signal is time varying.
all the previous sequence information is more effective than 2) L1 = 30, the frequency of the attack signal is 5 Hz, the
single-point analysis, while it will increase the computational amplitude of the attack signal is time varying.
HUANG et al.: FALSE DATA INJECTION ATTACK DETECTION FOR INDUSTRIAL CONTROL SYSTEMS 593

(a) (b) (c)

Fig. 10. Detection performance comparison among three methods. (a) Accuracy of three methods. (b) Sensitivity of three methods. (c) F1-score of three
methods.

TABLE II
D ETECTION P ERFORMANCE U NDER ATTACKS W ITH D IFFERENT F REQUENCIES

TABLE III
D ETECTION P ERFORMANCE U NDER ATTACKS W ITH D IFFERENT A MPLITUDES

The detection performance of the above two experiments are reaches 0.007 cm, which is 0.0035% of the original signal
shown in Tables II and III, respectively. amplitude, the detection performance is already good), the
As shown in Table II, as the frequency increases, the detection accuracy is over 98%. Because there may be occa-
accuracy and F1-score become higher, the sensitivity of sional disturbances or jitters during actual operation, it causes
μ-TF-HMM always keeps a higher level, and in most cases, a rare amount of false detections, which results in certain
the detection performance of μ-TF-HMM is better than μ- small fluctuations in the results. Through comprehensive com-
HMM. Similarly, as shown in Table III, as the amplitude parison, the detection performance of μ-TF-HMM is more
increases, the accuracy and F1 score also basically maintain stable and μ-TF-HMM detector maintains a good detection
an upward trend, the sensitivity of μ-TF-HMM always keeps performance.
a high level, and in most cases, the detection performance of
μ-TF-HMM is better than μ-HMM. Overall, as we can see
from the experimental results, μ-TF-HMM generally main- E. Real-Time Analysis
tains better detection performance than μ-HMM in the case Different from conventional information systems, the intru-
of most sinusoidal attacks. Especially when the amplitude is sion detection of the networked ICS requires a high real-
low, the difference in detection performance is more obvi- time performance. The proposed method includes the Fourier
ous. When the amplitude of the attack signal increases to a transform, feature extraction, and forward algorithm. The time
certain extent (e.g., when the amplitude of the attack signal complexities of the above operations are O(L2 ∗ log(L2 )),
594 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 1, JANUARY 1, 2021

TABLE IV
E XECUTION T IME U NDER ATTACKS W ITH D IFFERENT L ENGTHS OF S EQUENCES

O(L2 ), and O(L1 ∗ V 2 ), respectively. Since L1 , L2 , and V are [2] M. R. Khosravi and S. Samadi, “Reliable data aggregation in Internet
small, the amount of calculation of the proposed method is of ViSAR vehicles using chained dual-phase adaptive interpolation and
data embedding,” IEEE Internet Things J., vol. 7, no. 4, pp. 2603–2610,
small. To test the execution time of the three methods, 83 200 Apr. 2020.
samples are detected eight times in each method with differ- [3] A. Humayed, J. Lin, F. Li, and B. Luo, “Cyber-physical systems
ent lengths of sequences (CPU: i5-6400, System: Windows10, security—A survey,” IEEE Internet Things J., vol. 4, no. 6,
pp. 1802–1831, Jan. 2017.
GPU: GTX960). Finally, the average execution times are [4] D. Ding, Q.-L. Han, Y. Xiang, X. Ge, and X.-M. Zhang, “A survey
shown in Table IV. As the length of the sequence increases, on security control and attack detection for industrial cyber-physical
the execution time increases. The execution time of HMM is systems,” Neurocomputing, vol. 275, pp. 1674–1683, Jan. 2018.
the least, while the execution time of μ-HMM increases a [5] S. Kriaa, L. Pietre-Cambacedes, M. Bouissou, and Y. Halgand, “A sur-
vey of approaches combining safety and security for industrial control
little. The μ-TF-HMM spends the longest time and approxi- systems,” Rel. Eng. Syst. Safety, vol. 139, pp. 156–178, Jul. 2015.
mately equals to the double execution time of HMM. However, [6] A. Ameli, A. Hooshyar, A. H. Yazdavar, E. F. El-Saadany, and
the three methods all fully meet the real-time requirement for A. M. Youssef, “Attack detection for load frequency control systems
using stochastic unknown input estimators,” IEEE Trans. Inf. Forensics
the considered ICS. From the results in Tables III and IV, it Security, vol. 13, no. 10, pp. 2575–2590, Oct. 2018.
can be seen that μ-TF-HMM increases the execution time by [7] D. P. Fidler, “Was StuxNet an act of war? Decoding a cyberattack,”
80.32% (0.4874/0.2703) as compared with μ-HMM, while the IEEE Security Privacy, vol. 9, no. 4, pp. 56–59, Aug. 2011.
average F1-score increases 48.56% (0.9533/0.6417). Although [8] F. Meneghello, M. Calore, D. Zucchetto, M. Polese, and A. Zanella,
“IoT: Internet of Threats? A survey of practical security vulnerabilities in
μ-TF-HMM needs a little more execution time, it improves the real IoT devices,” IEEE Internet Things J., vol. 6, no. 5, pp. 8182–8201,
intrusion detection accuracy. In actual deployment, the appro- Jul./Aug. 2019.
priate L1 can be selected according to the computing power [9] A. Mukhopadhyay, S. Chatterjee, K. K. Bagchi, P. J. Kirs, and
G. K. Shukla, “Cyber risk assessment and mitigation (CRAM) frame-
of the equipment. work using logit and probit models for cyber insurance,” Inf. Syst. Front.,
Remark 1: The proposed intrusion detection algorithm can vol. 21, no. 5, pp. 997–1018, 2019.
be deployed in the controller node and sensor node, which [10] A. Matteini, F. Argenti, E. Salzano, and V. Cozzani, “A comparative
analysis of security risk assessment methodologies for the chemical
can effectively detect the FDI attacks from MITM attacks and industry,” Rel. Eng. Syst. Safety, vol. 191, Nov. 2019, Art. no. 106083.
internal attacks. [11] S. Aljawarneh, M. Aldwairi, and M. B. Yassein, “Anomaly-based
intrusion detection system through feature selection analysis and build-
V. C ONCLUSION ing hybrid efficient model,” J. Comput. Sci., vol. 25, pp. 152–160,
Mar. 2018.
This article investigated the intrusion detection problem for [12] C. Zhou et al., “Design and analysis of multimodel-based anomaly intru-
the ICS with repetitive machining under the FDI attacks. sion detection systems in industrial process automation,” IEEE Trans.
A data-driven intrusion detection method is proposed. The Syst., Man, Cybern., Syst., vol. 45, no. 10, pp. 1345–1360, Oct. 2015.
proposed method only utilizes the sensor measurements [13] M. Caselli, E. Zambon, and F. Kargl, “Sequence-aware intrusion detec-
tion in industrial control systems,” in Proc. 1st ACM Workshop Cyber
required in the ICS for closed-loop control and does not Phys. Syst. Security, 2015, pp. 13–24.
consume system resources additionally or require a precise [14] I. N. Fovino, A. Carcano, T. D. L. Murel, A. Trombetta, and M. Masera,
control system model. Besides, features in both time and “MODBUS/DNP3 state-based intrusion detection system,” in Proc. 24th
IEEE Int. Conf. Adv. Inf. Netw. Appl., 2010, pp. 729–736.
frequency domain are extracted at the same time. After feature [15] S. Ponomarev and T. Atkison, “Industrial control system network intru-
extraction, HMMs are established using features under normal sion detection by telemetry analysis,” IEEE Trans. Depend. Secure
operating conditions and applied for intrusion detection. The Comput., vol. 13, no. 2, pp. 252–260, Mar./Apr. 2015.
[16] B. Groza and P.-S. Murvay, “Efficient intrusion detection with bloom
experimental results verify that the proposed μ-TF-HMM has filtering in controller area networks,” IEEE Trans. Inf. Forensics Security,
the best intrusion detection performance as compared with vol. 14, no. 4, pp. 1037–1051, Apr. 2019.
HMM and μ-HMM methods. In the future, we will work [17] A. Almalawi, X. Yu, Z. Tari, A. Fahad, and I. Khalil, “An unsuper-
on the adaptive determination of thresholds, and more intelli- vised anomaly-based detection approach for integrity attacks on SCADA
systems,” Comput. Security, vol. 46, pp. 94–110, Oct. 2014.
gent decision-making methods to reduce human intervention [18] A. Khalili and A. Sami, “SysDetect: A systematic approach to critical
detection. state determination for industrial intrusion detection systems using a
priori algorithm,” J. Process Control, vol. 32, pp. 154–160, Aug. 2015.
[19] S. Zhanwei and L. Zenghui, “Abnormal detection method of industrial
R EFERENCES control system based on behavior model,” Comput. Security, vol. 84,
[1] X. Xu, C. He, Z. Xu, L. Qi, S. Wan, and M. Z. A. Bhuiyan, “Joint pp. 166–178, Jul. 2019.
optimization of offloading utility and privacy for edge computing [20] N. Erez and A. Wool, “Control variable classification, modeling and
enabled IoT,” IEEE Internet Things J., vol. 7, no. 4, pp. 2622–2629, anomaly detection in MODBUS/TCP SCADA systems,” Int. J. Crit.
Apr. 2020. Infrastruct. Protect., vol. 10, pp. 59–70, Sep. 2015.
HUANG et al.: FALSE DATA INJECTION ATTACK DETECTION FOR INDUSTRIAL CONTROL SYSTEMS 595

[21] M. Kalech, “Cyber-attack detection in SCADA systems using temporal [36] L. R. Rabiner, “A tutorial on hidden Markov models and selected appli-
pattern recognition techniques,” Comput. Security, vol. 84, pp. 225–238, cations in speech recognition,” Proc. IEEE, vol. 77, no. 2, pp. 257–286,
Jul. 2019. 1989.
[22] S. Adepu and A. Mathur, “Distributed attack detection in a water [37] Y. Qiao, X. Xin, Y. Bin, and S. Ge, “Anomaly intrusion detection method
treatment plant: Method and case study,” IEEE Trans. Depend. Secure based on HMM,” Electron. Lett., vol. 38, no. 13, pp. 663–664, 2002.
Comput., early access, Oct. 9, 2018, doi: 10.1109/TDSC.2018.2875008.
[23] J. Wang, D. Shi, Y. Li, J. Chen, H. Ding, and X. Duan, “Distributed
framework for detecting PMU data manipulation attacks with deep
autoencoders,” IEEE Trans. Smart Grid, vol. 10, no. 4, pp. 4401–4410,
Jul. 2019. Dajian Huang received the B.S. degree from
[24] X. Ying, S. U. Sagong, A. Clark, L. Bushnell, and R. Poovendran, the College of Information Engineering, Zhejiang
“Shape of the cloak: Formal analysis of clock skew-based intrusion University of Technology, Hangzhou, China, 2018,
detection system in controller area networks,” IEEE Trans. Inf. Forensics where he is currently pursuing the M.S. degree
Security, vol. 14, no. 9, pp. 2300–2314, Jul. 2019. in control theory and control engineering with the
[25] S. Ahmed, Y. Lee, H. Seung-Ho, and I. Koo, “Unsupervised machine College of Information Engineering.
learning-based detection of covert data integrity assault in smart grid His major research interests include cyber–
networks utilizing isolation forest,” IEEE Trans. Inf. Forensics Security, physical systems security and anomaly detection.
early access.
[26] M. Wan, W. Shang, and P. Zeng, “Double behavior characteris-
tics for one-class classification anomaly detection in networked con-
trol systems,” IEEE Trans. Inf. Forensics Security, vol. 12, no. 12,
pp. 3011–3023, Dec. 2017.
[27] C. Shen, C. Liu, H. Tan, Z. Wang, D. Xu, and X. Su, “Hybrid- Xiufang Shi (Member, IEEE) received the B.Sc.
augmented device fingerprinting for intrusion detection in industrial degree in automation from the East China University
control system networks,” IEEE Wireless Commun., vol. 25, no. 6, of Science and Technology, Shanghai, China, in
pp. 26–31, Dec. 2018. 2011, and the Ph.D. degree in control science and
[28] W. Xu, C. Yan, W. Jia, X. Ji, and J. Liu, “Analyzing and enhancing the engineering from Zhejiang University, Hangzhou,
security of ultrasonic sensors for autonomous vehicles,” IEEE Internet China, in 2016.
Things J., vol. 5, no. 6, pp. 5015–5029, Dec. 2018. She was a joint Ph.D. student with the University
[29] B. Kailkhura, Y. S. Han, S. Brahma, and P. K. Varshney, “Distributed of Sydney, Sydney, NSW, Australia, in 2015.
Bayesian detection in the presence of Byzantine data,” IEEE Trans. From 2016 to 2018, she was a Postdoctoral
Signal Process., vol. 63, no. 19, pp. 5250–5263, Oct. 2015. Researcher with the College of Control Science
[30] B. Kailkhura, Y. S. Han, S. Brahma, and P. K. Varshney, “Asymptotic and Engineering, Zhejiang University. She is cur-
analysis of distributed Bayesian detection with Byzantine data,” IEEE rently an Associate Professor with the Zhejiang University of Technology,
Signal Process. Lett., vol. 22, no. 5, pp. 608–612, May 2015. Hangzhou. Her major research interests include industrial control system
[31] Y. Huang, J. Tang, Y. Cheng, H. Li, K. A. Campbell, and Z. Han, security, wireless localization, and target tracking.
“Real-time detection of false data injection in smart grid networks: An
adaptive CUSUM method and analysis,” IEEE Syst. J., vol. 10, no. 2,
pp. 532–543, Jun. 2016.
[32] K. Manandhar, X. Cao, F. Hu, and Y. Liu, “Detection of faults and Wen-An Zhang (Member, IEEE) received the
attacks including false data injection attack in smart grid using Kalman B.Eng. degree in automation and the Ph.D. degree
filter,” IEEE Trans. Control Netw. Syst., vol. 1, no. 4, pp. 370–379, in control theory and control engineering from
Dec. 2014. the Zhejiang University of Technology, Hangzhou,
[33] Y. Guan and X. Ge, “Distributed attack detection and secure estimation China, in 2004 and 2010, respectively.
of networked cyber-physical systems against false data injection attacks He has been with the Zhejiang University of
and jamming attacks,” IEEE Trans. Signal Inf. Process. Netw., vol. 4, Technology since 2010, where he is currently
no. 1, pp. 48–59, Mar. 2018. a Professor with the Department of Automation.
[34] S. Amin, X. Litrico, S. Sastry, and A. M. Bayen, “Cyber security of He was a Senior Research Associate with the
water SCADA systems—Part I: Analysis and experimentation of stealthy Department of Manufacturing Engineering and
deception attacks,” IEEE Trans. Control Syst. Technol., vol. 21, no. 5, Engineering Management, City University of Hong
pp. 1963–1970, Sep. 2012. Kong, Hong Kong, from 2010 to 2011. His current research interests include
[35] S. Amin, X. Litrico, S. S. Sastry, and A. M. Bayen, “Cyber security cyber–physical systems security, networked control systems, and multisensor
of water SCADA systems—Part II: Attack detection using enhanced information fusion estimation
hydrodynamic models,” IEEE Trans. Control Syst. Technol., vol. 21, Prof. Zhang was awarded an Alexander von Humboldt Fellowship in
no. 5, pp. 1679–1693, Sep. 2012. 2011–2012.

You might also like