Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

Robust Multivariate Anomaly Based Intrusion Detection System for Cyber-Physical

Systems using Low Rank Matrix

Aneet K. Dutta Bhaskar Mukhoty Sandeep K. Shukla


C3i Center,Computer Science & C3i Center,Computer Science & C3i Center,Computer Science &
Engineering Department Engineering Department Engineering Department
Indian Institute of Technology Kanpur Indian Institute of Technology Kanpur Indian Institute of Technology Kanpur
India India India
Email: aneet@cse.iitk.ac.in Email: bhaskarm@cse.iitk.ac.in Email: sandeeps@cse.iitk.ac.in

Abstract—Motivation, problem statement, approach, re- There are two kinds of approaches in designing an IDS,
sults,conclusions model-driven approach and data-driven approach. In model-
The driven based IDS consists of rules which check whether
any system dynamics are not violating. The IDS in Indus-
1. Introduction trial Control System(ICS) generally models the behavior
of the system because of the reason that the system fol-
The entire economy of a nation fuels by the proper lows the proper law of physics. These systems are large
functioning of the critical infrastructure. The critical infras- and complex, consisting of many physical parameters that
tructures include health care, manufacturing, water treat- are measured by devices called sensors and controlled by
ment, transportation, and power system. The disruption in actuators. Therefore, manually identifying rules to model
the service of critical infrastructures can affect people’s lives the behavior of these complex systems is not feasible. The
even for a minimal amount of time. With the advancement of data-driven approach helps us in sketching ICS’s behavior
technology in network communications, the SCADA, an es- and understanding the system’s underlying dynamics. In
sential component of any critical infrastructures, has moved real-time, when the system’s functioning deviates from the
to IP based communication from serial communication lead- modeled behavior beyond a predetermined threshold, it is
ing to an increase in the attack surface for the attackers considered an anomaly.
to penetrate these systems. With the IP convergence, their With advancements in the field of machine learning and
networks are further integrated with the IT network and deep learning, many data-driven based IDS exists. Since
exposed to adversaries outside the operational technologies this machine learning and deep learning-based intrusion
(OT) networks. There is an immediate need to detect the detection methods are dependent on the data, there are some
attacks in these systems and generate an alert in real-time. challenges and concerns. These are:
For this reason for detecting the attacks in real-time, the
• the unavailability of properly labeled data of the
Intrusion Detection System must be of low time complexity
cyber-physical system’s behavior.
during the test phase.
• Most machine learning and deep learning-based IDS
The OT engineers ensured the proper functioning of
are not outlier resistant and adversarial resistant.
this cyber-physical system during the design phase without
Thus, the robustness of these detection mechanisms
considering the security threats to this system. Consequently,
is not guaranteed.
many security vulnerabilities exist in the cyber-physical
• The use of highly sophisticated machine learning
systems, like in the industrial communication protocol like
or deep learning models may not be feasible in a
MODBUS. The packets are unencrypted and do not have any
resource-constrained environment.
authorization mechanism to validate the sender, thus making
the protocol vulnerable to man in the middle (MITM) attack In this paper, our work focuses on devising a robust
and false data injection. The vulnerabilities like remote code anomaly detection algorithm that will learn the deterministic
execution(RCE), improper credential management, stack & structure of the cyber-physical system and detect anomaly
buffer overflow, and memory corruption make the SCADA when the system’s behavior deviates from the modeled be-
software and HMI devices easy targets for the attackers. havior learned by the algorithm. The deterministic structure
Now, changing all the pieces of equipment in existing of the cyber-physical system’s behavior is obtained in a
infrastructure with updated and secured protocols are time- low dimensional subspace by recovering a low-rank ma-
consuming and expensive. Instead, we developed a solution trix by applying the Robust Principal Component Analysis
that can retrofit with the existing system to protect the technique. The orthonormal basis vectors of these low-
system with the detection of attacks in real-time. dimensional subspaces are determined, which will help us
determine the closest point of the given vector comprising the different variables(sensors/actuators). To generate the
of readings from sensors and actuators associated with the correct invariants, the intrusion detection developer must
system in the low dimensional subspace. A cluster forms have detailed knowledge about the cyber-physical system
of the vectors projected in the low dimensional space, and dynamics.
a threshold is determined. At the test phase, vectors are Bernabeu et al. [1] developed an intrusion detection
projected in the low-dimensional subspace, and based on the mechanism based on a decision tree algorithm to protect the
distance of the test vector to the cluster anomaly is flagged. power grid. [1] demonstrates an interesting fact about how
The main contribution of this paper is: the critical points in the power-grid changes during differ-
ent environmental scenarios and the necessity of different
• We propose a multivariate model to detect intrusion detection models in different seasons. Hence, the
anomalies in the behavior of the cyber physical supervised learning based intrusion detection mechanisms
systems with limited resource requirements. need to be updated with properly labeled data of normal and
• The proposed intrusion detection model is ro- attack classes at different environmental conditions. How-
bust and outlier resistant. Also it is adversarial ever, the availability of adequately labeled data of cyber-
resistant(Poisoning attack in which the adversary physical systems at different conditions is a considerable
contaminates the training data). challenge.
• Our method retrofits to the existing infrastruc- Since the availability of adequately labeled data of the
ture of an ICS. cyber-physical system’s behavior is a challenge, various
• The IDS can be developed with very small anomaly-based intrusion detection systems are proposed.
amount of labeled data(only labeled as Normal The idea behind any anomaly-based intrusion detection
behavior). mechanism is that the cyber-physical system’s normal be-
• The IDS developed requires very limited memory havior is modeled, and the deviation from the modeled
and is time efficient making it suitable to deploy behavior is measured during the run-time. If the measure
for detecting anomalies in real-time. of the deviation crosses the pre-determined threshold, the
anomaly is flagged.In the papers [7], [8], [14] and [15]
2. Background and Related Work anomaly-based intrusion detection techniques are proposed.
In this paper, RNN and LSTM are used to build a predictive
With the rapid progress towards industry 4.0 and IP- model for each variable individually by training using the
convergence the risk of cyberattacks in the cyber-physical system’s normal behavior. At each time step, the model
systems are increasing every day. The chances of state- predicts the next reading. The difference between the actual
sponsored attacks on the cyber-physical systems are high reading and the predicted reading is measured, and if the dif-
and proper detection, and prevention mechanism needs ference is above the predetermined threshold, then anomaly
to be in place to protect this infrastructures. There has is flagged. The underlying assumption for developing these
been a significant amount of research going on in this anomaly-based IDS is that the data used for training the
area. Researchers from different backgrounds have pro- models is outlier free. These may not be the case in a
posed intrusion detection mechanisms for cyber-physical practical scenario due to the malfunctioning of different
systems.Goldenberg et al. [2] proposes a deterministic fi- sensors resulting in corruption or missing data in the training
nite automata(DFA) based intrusion detection system. The dataset. The presence of outliers or corrupted data in the
network traffic of the cyber-physical system is modeled training set will profoundly affect the IDS’s performance
by constructing a DFA. The construction of minimal-sized and accuracy. Aoudi et al. [10] proposed an anomaly-based
DFA will reach a final state when the system is under IDS in which the normal behavior of the cyber-physical
normal behavior and will reach a non-final or dead state system is modeled, and during the test phase, the system’s
when the system behaves maliciously is NP-hard. DFA’s can current behavior is measured to the modeled behavior of the
be constructed for an instance of cyber-physical systems, system. If the deviation between the current behavior and the
but DFA based IDS are highly sensitive and have a high modeled behavior crosses a certain threshold, an anomaly is
rate of false-positives.Cheng et al. [11] proposed a similar flagged. For each variable, the time-series data is converted
idea to [1] which is an intrusion detection mechanisms to a Hankel matrix. A low-dimensional subspace is identified
using finite-state automata (FSA). FSA is designed to model that captures the deterministic structure of the cyber-physical
the programs’ control flow, i.e., system calls and memory system’s dynamics with the help of SVD. Since the SVD is
addresses referred to or API calls executed in the cyber- applied to the covariance matrix XX T the IDS proposed is
physical system. The IDS proposed by Cheng et al. has not robust to the outliers present in the training data. Also,
significant overhead and time constraints because of con- this proposed IDS are univariate and separate models are
tinuous monitoring of the CPS software’s system calls. It need to be trained and deployed in the system for detecting
has a limited ability to detect attacks as it only detects the attacks in the overall infrastructure, it is not feasible for
when the attacker changes the execution control flow in deployment in resource constrained environments.
the cyber-physical system. [5] propose an invariants based The ideas in the papers [6] and [13] are similar to these
technique for generating alarms when the system behavior in [7], [8], [14] and [15] but in a multivariate setting. The
violates the invariants derived from the correlations between multivariate IDS solves the problem of deploying separate
models for each variable and can be deployed in real- the differential pressure between the two tanks. Therefore,
time in a resource-constrained environment. However, the this deterministic structure in the behavior of the cyber-
underlying assumption of outlier free training data and the physical system is mathematically modeled. The system’s
whole training data belongs to the normal class persists. malicious or anomalous behavior is detected when the on-
Therefore, these IDS mechanisms can not guarantee robust- going behavior deviates from the system’s mathematically
ness essential for a practical scenario. modeled behavior. Thus, intrusion detection in the cyber-
physical system is a general anomaly detection problem.
3. System Architecture The deep learning and machine learning techniques are
widely used to solve the anomaly detection problem. How-
In our work, the communication link between the PLC ever, specific security issues exist in these techniques, which
and SCADA is the threat vector due to vulnerable industrial will enable the attacker to create carefully crafted inputs
communication protocols like MODBUS. Since command such that these intrusion detection mechanisms will fail to
injection, false data injection, replay, and MITM attacks are classify these examples as an anomaly correctly. Therefore,
possible due to MODBUS’s vulnerability, it will enable the the robustness of the intrusion detection mechanisms is
attacker to change the dynamics of the system. The attack essential to protect the model against outliers, missing data,
scenarios generated are by manipulating the sensor and ac- and poisoning of the training data.
tuator values in the data packets communicated between the In this paper, our work focuses on devising a robust
PLC and SCADA (false data injection attacks). Any number anomaly detection algorithm that will learn the deterministic
of sensors or actuators or combinations of any number of structure of the cyber-physical system and detect anomaly
sensors or actuators can be modified, which depicts that the when the system’s behavior deviates from the modeled
number of attack scenarios or the number of undesired state behavior learned by the algorithm.
of the cyber-physical system is infinite. These are the following steps to develop the Intrusion
Detection System.
The Intrusion Detection System is deployed within the
SCADA host, which collects the infrastructure’s entire data • Dataset Collection
and maintains the overall state. IDS read the measurements • Anomaly Detection Algorithm
of the sensors and actuators from the historian database to • Evaluation of the the proposed methodology in the
determine that the cyber-physical system’s current behavior Test Phase.
is normal or anomalous. • Time & Space Analysis
• The performance of the proposed IDS in the pres-
ence of the outliers.
• Deployment of the IDS in real-time.

Figure 1. Architecture of the integrated SCADA software & IDS within


SCADA host

Figure 1 describes how the data flows between the dif-


4.1. Dataset Collection
ferent components within the SCADA host. The readings of
4.1.1. Secure Water Treatement(SWaT) Testbed. In the
all the sensors and actuators, along with the timestamp, are 51
SWaT dataset {xi , yi }N
i=1 where xi IR , N is the number
logged in to the historian server. The SCADA software reads
of datapoints (record) and yi is the label depicting with
the data from the historian server to monitor the system’s
Normal/Attack. The vector xi consists of reading of ac-
current readings and state. Our proposed IDS deployed in
tuators and sensors at each timestamp. In SWaT dataset,
the SCADA host reads the data from the historian server
N = 449919, out of which 395298 are normal data points
in real-time, calculates the anomaly score, and sends the
and 54621 are attack data points comprising of 36 attack
necessary information to the notification system, and stores
scenarios.
it in the event log database.
The Intrusion Detection System(IDS) developed by con-
4.1.2. Tennessee-Eastman Process. In the TE process
sidering the system’s behavioral characteristics will help to 41
dataset {xi }N
i=1 where xi IR , N is the number of data-
defend the infrastructure against the attackers who are trying
points (record). The vector xi consists readings of sensors
to change the state of the system to an undesirable state by
at each timestamp. In TE process dataset, N = 4801, out of
manipulating the dynamics of the system.
which first 4000 data points belongs to normal behavior of
the system and the last 800 are observations corresponding
4. Proposed Methodology to anomalous behavior of the system.

The behavior of the cyber-physical systems has a deter- 4.1.3. Govt. Sponsored Power Testbed in India. In the
31
ministic structure following the laws of Physics. Like for power testbed dataset {xi }Ni=1 where xi IR , N is the
simple example, if the heater is on, the temperature will number of datapoints (record). The vector xi consists read-
increase, or the inflow rate of the water is determined from ings of sensors corresponding to voltage, current, power
and phase angle reading at each timestamp. In this dataset, {b1 , b2 , b3 , ........, bk } obtained from the low rank matrix L.
N = 12310, out of which first 12125 data points belongs The basis vectors {b1 , b2 , b3 , ........, bk } will span the low
to normal behavior of the system and 185 are observations dimensional subspace S , which captures the deterministic
corresponding to the behavior of the system when the system structure of the cyber-physical system’s behavior.
is subjected under attack. The basis vectors {b1 , b2 , b3 , ........, bk } must be or-
thonormal to each other to make the computation faster
4.2. Anomaly Detection Algorithm in the testing phase which will allow us to deploy the
intrusion detection system in resource constraint devices. If
The first phase of devising the anomaly detection algo- the basis vectors {b1 , b2 , b3 , ........, bk } are not orthonormal
rithm is to extract the deterministic structure of the cyber- to each other, then Gram–Schmidt method is applied to
physical system’s behavior. orthonormalize this set of vectors.
Therefore, V is the set of orthonormal basis vectors
4.2.1. Determining Deterministic Structure of the Cyber- {v1 , v2 , v3 , ...., vk } derived from the low rank matrix L or
Physical System’s behavior–Robust Principal Compo- by applying Gram-Schmidt method to B . This orthonormal
nent Analysis. Since the cyber-physical system’s behavior basis vectors are organized as a matrix A in the form of
has a deterministic structure, the data will lie in a low column vectors.
dimensional subspace. PCA is a widely used technique for
dimensionality reduction, but it is highly brittle and outlier We know, xIRn which can be expressed as x = v + w
sensitive. The low dimensional subspace of the data is where v is the projection of x on the subspace S and w is
determined by using Robust PCA(RPCA), which makes the the orthogonal complement.
intrusion detection mechanism resistant to outliers, missing
data, and against poisoning of the training data. P rojv x = A(AT A)−1 AT x
The two matrices L(low-rank matrix) and S0 (Sparse Since, {v1 , v2 , v3 , ...., vk } are orthonormal to each other,
matrix) are recovered given a large data matrix M where L therefore (AT A) = Ik .
is the low rank matrix representing the deterministic stricture P rojv x = AAT x
of the system’s behavior in low-dimensional subspace and
S0 is the sparse matrix containing the outliers. This problem The euclidean norm of the ||P rojv x||2 is given by:
is solved by a convex program called Principal Component
Pursuit which minimizes a weighted combination of the
nuclear norm and l1 norm. Solving this problem suggests ||AAT x||2 = (AAT x).(AAT x)
that we will be able to recover the principal components of
the data matrix even if some of the entries in the matrix are ||AAT x||2 = (AAT x)T (AAT x)
corrupt or outliers.
||AAT x||2 = xT A(AT A)AT x
Optimization Problem
Since, (AT A) = I
The optimization problem to solve RPCA problem is:
minimize ||L||∗ + λ||S0 ||1 ||AAT x||2 = xT AAT x
subject to L + S = M
Solving the above optimization problem exactly recovers ||AAT x||2 = (Ax)T (Ax)
low-rank L and sparse S0 because:
||AAT x||2 = ||AT x||2
• The nuclear norm of the L matrix is the sum of the
singular values of L. Hence, it is proved that the square euclidean norm
• Since the number of non-zero singular values is the ||AAT x|| = ||AT x||.
rank of a matrix, therefore minimizing the nuclear
norm will yield us a low-rank matrix.
After obtaining A, xi is projected in this low-
• Minimizing the l1 norm of the matrix S will ensure
dimensional subspace S and the euclidean norm of it is
sparsity of S0 .
calculated, where vector xi consists of reading actuators
The convex PCP problem using Augmented Lagrangian and sensors at each timestamp. The xi ’s are projected in
Multiplier(ALM). the low-dimensional subspace because the P rojv xi is the
closest point of xi in the subspace S . The median of the
4.2.2. Mathematical Representation of the Deterministic projected points in subspace S is calculated.
Structure of the Cyber-Physical System’s behavior.
After obtaining the deterministic structure of the cyber- Let m be the number of data points projected in the
physical system’s behavior, the mathematical representation subspace S
of this behavior is determined. B is the set of basis vectors
For each xi in the training dataset, Zi = AT xi , where The time complexity of this matrix multiplication is
Z is a matrix storing all the projected datapoint xi in the O(rd).
subspace S . • Euclidean distance is calculated between two r ∗ 1
med=median of Z along axis-X. vector. The time complexity of calculating the Eu-
The deterministic behavior of the cyber-physical system clidean norm between this two vectors is O(r).
can be mathematically represented by the matrix A and the • The calculated euclidean norm is compared with the
vector med. threshold(Θ) in O(1) time.
Therefore, the time complexity of our proposed IDS is
4.2.3. Threshold Determination(Θ). The median of the O(rd).
distribution is the center of probability. It is the point Since, r << d, the proposed IDS functions in O(d)
situated at the center of the sample where the probability space and time complexity.
is exact 21 .
5. Experiments & Results
The threshold(Θ) depicts the maximum deviation of the
cyber-physical system’s behavior from the median which is To evaluate the performance and accuracy of the pro-
allowed by the anomaly detection algorithm. The value of posed intrusion-detection model, we have used the SWaT
the threshold(Θ) is determined by calculating the highest dataset, Tennessee Eastman process dataset and a dataset
anomaly score recorded when the algorithm is run for a corresponding to a Government of India sponsored power
set of data points which are already classified as Normal distribution testbed.
behavior.
5.1. Training Phase
T 2
AnomalyScore(d) = ||med − A xi || 5.1.1. Secure Water Treatement(SWaT) Testbed. The
data matrix M of dimension (496800 ∗ 51) is given as an
Θ = max{di } where i[1, n] where di is the anomaly input to the RPCA algorithm to recover the low rank matrix
score corresponding to each data point in the which is L. The rank r of L is found to be 18. Figure ?? show
pre-classified as normal behavior. The number of data how the error which is the Frobenius norm of ||M − L − S||
points in the set is n. due to the constraint mentioned in the optimization problem
L + S = M , reduces at each epoch.. The stopping criteria is
Therefore, the deterministic or normal behavior of either 10, 000 epochs or when error is less than 10−7 ||M ||F .
the cyber-physical system can be mathematically repre- For our intrusion detection model, the number of epochs
sented by the matrix A, the vector med and the scalar required for obtaining the low-rank matrix L by applying
value Θ. RPCA is 5970.

4.3. Classification of the Test Behavior

The IDS classifies by calculating the anomaly score(di )


and comparing it with the threshold(Θ). If the anomaly
score(di ) is higher than the threshold(Θ), then the data point
xi is flagged as an anomaly.

4.4. Time & Space Analysis

The space required for deploying the the intrusion de-


tection model is:
• Matrix A of dimension d ∗ r; where d is the number
features in matrix M and r is the rank of low rank Figure 2. Error calculated at each Epochs while running RPCA on SWaT
Dataset
matrix L.
• A vector of size r. The orthonormal basis vectors {v1 , v2 , v3 , ...., v18 } are
• A scalar value Θ. each of (51 ∗ 1) dimension. The matrix A comprising of the
Therefore, the space complexity of our proposed IDS is orthonormal basis vectors of L is of dimension (51 ∗ 18).
O(rd). The datapoints xi each of (51∗1) dimension is projected on
The steps involved in classifying a test data point are as the subspace S by AT xi resulting in the dimension of each
follows: projected vector to be of the dimension (18∗1). The median
vector from the set of all projected vectors is calculated. The
• A matrix multiplication between AT and xi where median is a (18 ∗ 1) column vector and the threshold(Θ) is
AT is r ∗ d dimension and xi is d ∗ 1 dimension. a scalar value.
5.1.2. Tennessee Eastman Process Dataset. The data ma- 5.2. Test Phase
trix M is of dimension (3000∗41) is given as an input to the
RPCA algorithm to recover the low rank matrix L. The rank 5.2.1. Secure Water Treatement(SWaT) Testbed. After
r of L is found to be 24. Figure 3 show how the error which the training phase there are three learning parameters that are
is the Frobenius norm of ||M − L − S|| due to the constraint stored i.e, A of (51∗18) dimension, median vector of (18∗1)
mentioned in the optimization problem L+S = M , reduces dimension and a scalar value Θ for the functioning of the
at each epoch. For our intrusion detection model, the number intrusion detection model in test phase. This shows that the
of epochs required for obtaining the low-rank matrix L by intrusion detection model designed by our algorithm does
applying RPCA is 4269. not require much memory and can be stored in a resource
constrained environment. For demonstration, we have shown
how our IDS detects the ongoing attack considering four
different attack scenarios.

Attack Scenario 1. In attack scenario 1, the attacker’s intent


is to overflow the tank 1 by attacking on the actuator MV-
101 to remain open even if the level indicator sensor LIT-
101 is above the highest permitted level. Figure 5 shows the
reading of the sensor LIT-101 at each time step. The highest
permitted level is 800mm. We can see that the attacker is
successfully overflowing the tank from the time step 1400
to 2600. Then, when the attack stopped the level of tank
comes within the permitted level.

Figure 3. Error calculated at each Epochs while running RPCA on TE


Process Dataset

The matrix A comprising of the orthonormal basis vec-


tors of L is of dimension (41 ∗ 24). The median is a (24 ∗ 1)
column vector.

5.1.3. Power Distribution Dataset. The data matrix M is


of dimension (3000 ∗ 31) is given as an input to the RPCA
algorithm to recover the low rank matrix L. The rank r of
L is found to be 9. Figure 4 show how the error which is
the Frobenius norm of ||M − L − S|| due to the constraint
mentioned in the optimization problem L+S = M , reduces
at each epoch. For our intrusion detection model, the number Figure 5. LIT-101 Sensor reading at each Time Step
of epochs required for obtaining the low-rank matrix L by
applying RPCA is 9269.

Figure 6. Anomaly Score at each Time Step

Figure 6 shows the anomaly score at each time step with


Figure 4. Error calculated at each Epochs while running RPCA on Power
Dataset
the blue horizontal line as the threshold. It clearly shows
that when the attack started and eventually when the level of
The matrix A comprising of the orthonormal basis vec- water is above the permitted level the anomaly score crosses
tors of L is of dimension (31 ∗ 9). The median is a (9 ∗ 1) the threshold flagging this behavior of the cyber-physical
column vector. system as an anomaly. Again, when the attack stopped and
the level of the water in the tank is within the permitted level in the sensor reading describes that the intrusion detection
the anomaly score decreases and comes below the threshold model can capture the deviation in system’s behavior.
classifying the system’s behavior as normal.
5.2.2. Tennessee Eastman Process Dataset. After the
Attack Scenario 2. In attack scenario 2, the attacker’s training phase there are three learning parameters that are
intent is to overflow the tank 3 by attacking on the sensor stored i.e, A of (41∗24) dimension, median vector of (24∗1)
LIT-301 by increasing the level of the tank by 1mm at dimension and a scalar value Θ for the functioning of the
each second. The attacker did not stop and the tank does intrusion detection model in test phase.
overflow(> 800mm) for a period of time. 7 shows the
reading of the sensor LIT-301 at each time step. The highest
permitted level is 800. We can see that the attacker is Attack Scenario 1. In attack scenario 1, the attacker’s intent
successfully overflowing the tank. Then, when the attack is to increase the value of the sensor abruptly by gradually
stopped the level of tank comes within the permitted level. increasing to a point well above the desired level. Figure 9
shows the reading of the sensor XMeas(14) with time when
the sensor is under attack.

Figure 7. LIT-301 Sensor reading at each Time Step Figure 9. Xmeas(14) Sensor reading at each Time Step

Figure 10. Anomaly Score at each Time Step

Figure 8. Anomaly Score at each Time Step

Figure 8 shows the anomaly score at each time step with Attack Scenario 2. In attack scenario 2, the attacker’s intent
the blue horizontal line as the threshold. It clearly shows is to increase the value of the sensor abruptly by gradually
that when the attack started and eventually when the level increasing to a point well above the desired level. Figure 11
of water is above the permitted level the anomaly score shows the reading of the sensor XMeas(6) with time when
crosses the threshold explained by the peak like structure in the sensor is under attack.
the plot flagging this behavior of the cyber-physical system Figure 10 and Figure 12 shows the anomaly score at each
as an anomaly. Again, when the attack stopped and the level time step with the blue horizontal line as the threshold. It
of the water in the tank is within the permitted level the clearly shows that when the attack started and eventually
anomaly score decreases and comes below the threshold when the sensor reading is above the permitted level the
classifying the system’s behavior as normal. The sharp drop anomaly score crosses the threshold flagging this behavior
in the anomaly score plot corresponding to the sharp drop of the cyber-physical system as an anomaly.
Figure 11. Xmeas(6) Sensor reading at each Time Step Figure 14. Anomaly Score at each Time Step

ior of the cyber-physical system, the spike in the anomaly


score plot depicts that that this abnormality in the behavior
is detected.

5.2.4. Robustness in the Presence of Outliers in the


Training Data. There are two kinds of outliers injected
into the training data:

• Gaussian Noise: Gaussian Noise with mean(µ) 0


and variance(σ) 1 is introduced to the whole train-
Figure 12. Anomaly Score at each Time Step ing data.
• Burst Outliers: Some variables are targeted and the
data of that variables are corrupted over for a con-
5.2.3. Power Distribution Dataset. After the training phase secutive time stamps of length(l) at regular interval.
there are three learning parameters that are stored i.e, A of
(31∗9) dimension, median vector of (9∗1) dimension and a Tennessee Eastman Process Dataset. The variables
scalar value Θ for the functioning of the intrusion detection XMeas(6) and XMeas(14) are targeted, and the data of that
model in test phase. variables are corrupted at a regular interval of 25 time steps.
In this attack scenario, the attacker’s intent is to increase Figure 15 shows the reading of a variable(Xmeas(6))
the value of the sensor ’PT transcontractor C3’ which is a which is targeted and outliers are injected at regular in-
measures transconductance in the third phase of a 3-phase tervals. The spikes in the reading depicts the presence of
power distribution testbed suddenly generating a spike in the outliers in the training data.
reading.Figure 13 shows the reading of the sensor XMeas(6)
with time when the sensor is under attack.

Figure 15. Sensor data reading in the presence of outliers

Figure 13. PT transcontractor C3 Sensor reading at each Time Step Figure 16 shows the similarity with Figure 12 and is able
which shows the robustness of the proposed IDS to detect
Figure 14 shows the anomaly score at each time step the anomalies in the test phase though outliers (in the form
with the blue horizontal line as the threshold. It clearly of Gaussian noise or random ourliers) were present in the
shows that when there is a sudden abnormality in the behav- training data.
[3] Yoshihiro Hashimoto,Takeshi Toyoshima, Shuichi Yogo, Masato
Koike,Takashi Hamaguchi,Sun Jing and Ichiro Koshijima,”Safety
securing approach against cyber-attacks for process control sys-
tem”,Computers and Chemical Engineering 57 (2013) 181– 186.
[4] Youbiao He, Gihan J.Mendis and Jin Wei,”Real-Time Detection of
False Data Injection Attacks in Smart Grid: A Deep Learning-Based
Intelligent Mechanism”,IEEE TRANSACTIONS ON SMART GRID,
VOL. 8, NO. 5, SEPTEMBER 2017.
[5] Sridhar Adepu and Aditya Mathur,”Distributed Attack Detection in a
Water Treatment Plant: Method and Case Study”,2018.
[6] Dan Li,Dacheng Chen,Lei Shi,Baihong Jin,Jonathan Goh and See-
Kiong Ng,”MAD-GAN: Multivariate Anomaly Detection for Time Se-
ries Data with Generative Adversarial Networks”,arXiv:1901.04997v1
Figure 16. Anomaly Score at each Time Step [cs.LG] 15 Jan 2019.
[7] J.Goh,S.Adepu, M.Tan and Z.S.Lee, “Anomaly detection in cyber-
physical systems using recurrent neural networks”, inProc. of
The rank r of the recovered low rank matrix L is 24 HASE.IEEE, 2017, pp. 140–145.
with out the injection of outliers in the training data, and [8] P.Malhotra,L.Vig,G.Shroff and P.Agarwal,“Long short term memory
rank r of the recovered low rank matrix L is 23 with the networks for anomaly detection in time series”, inProc. ofESANN,
2015, p. 89.
injection of outliers in the training data.
[9] J.Goh,S.Adepu,K.N.Junejo,and A. Mathur, “A dataset to support re-
search in the design of secure water treatment systems,” inProc.
5.2.5. Deployment of IDS in Real-time and the Memory ofCRITIS, 2016.
Requirement. Considering the dataset of the power dis-
[10] Wissam Aoudi,Mikel Iturbe,and Magnus Almgren,”Truth Will Out:
tribution testbed, where A is a (31 ∗ 9) matrix and med Departure-Based Process-Level Detection of Stealthy Attacks on Con-
is a 9 dimensional vector and Θ is floating point vari- trol Systems”,CCS’18, October 15-19, 2018, Toronto, ON,Canada.
able. The memory required to store the parameters required [11] Long Cheng,Ke Tian,Danfeng(Daphne) Yao, Lui Sha,and Raheem A.
for calculating the anomaly score and comparing with the Beyah,”Checking is Believing: Event-Aware ProgramAnomaly Detec-
threshold value to decide whether to generate alarms or not tion in Cyber-Physical Systems”,arXiv:1805.00074v2[cs.CR] 25 Mar
is (31 ∗ 9 ∗ 8) + (9 ∗ 8) + 8 = 2312 bytes i.e., 2.312KB. 2019.
The general-purpose machines in which the SCADA [12] T. Morita, S. Yogo, M. Koike, T. Hamaguchi, S. Jung, I. Koshijima,
software operates have a RAM of 256MB and running and Y. Hashimoto,”Detection of Cyber-Attacks with Zone Dividing
and PCA”,in the proceedings of the 17th International Conference on
simple code for matrix multiplication, and a conditional Knowledge based and Intelligent Information and Engineering Systems
operation is feasible. Therefore, the proposed IDS is feasible 2013.
to be deployed in real infrastructure within the SCADA host [13] Pankaj Malhotra, Anusha Ramakrishnan, Gaurangi Anand, Lovekesh
to detect attacks in real-time. Vig, Puneet Agarwal, Gautam Shroff,”LSTM-based Encoder-Decoder
Generally, Programmable Logic Controllers(PLC) have for Multi-sensor Anomaly Detection”,ICML 2016 Anomaly Detection
Internal RAM of 256kB. The memory required to store Workshop, NewYork, NY, USA, 2016.
the IDS model is 2.312KB. If matrix multiplication and [14] Pavel Filonov, Fedor Kitashov, and Andrey Lavrentyev,”RNN-
comparison of two variables are allowed in the PLC, which based Early Cyber-Attack Detection for the Tennessee Eastman Pro-
cess”,arXiv:1709.02232v1 [cs.CR] 7 Sep 2017.
is generally possible using STL language, then our proposed
IDS can be deployed within PLC. [15] Pavel Filonov,Andrey Lavrenty, and Artem Vorontsov,”Multivariate
Industrial Time Series with Cyber-AttackSimulation: Fault
Detection Using an LSTM-based Predictive Data Model”,
6. Conclusion arXiv:1612.06676v2cs.LG] 26 Dec 2016.
[16] C.Alcaraz, G.Fernandez,and F.Carvajal,“Security aspects of SCADA
The conclusion goes here. and DCS environments”, inCritical Infrastructure Protection, pp.
120–149, Springer, 2012.

Acknowledgments [17] G.Y.Liao, Y.J.Chen, W.C.Lu, and T.C.Cheng,“Toward authenticating


the master in the modbus protocol”,IEEE Transactions on PowerDe-
livery, vol. 23, no. 4, pp. 2628–2629, 2008.
The authors would like to thank... [18] US National Institute of Standards and Technology,”NIST cyber
security framework”, USA 2018.
References [19] Emmanuel J. Candes, Xiaodong Li, Yi Ma, and John Wright,”Robust
Principal Component Analysis?”, arXiv:0912.3599v1 [cs.IT] 18 Dec
[1] Emanuel E.Bernabeu,James S.Thorp and Virgilio Cen- 2009.
teno,”Methodology for a Security/Dependability Adaptive Protection
Scheme Based on Data Mining”,IEEE TRANSACTIONS ON POWER
DELIVERY, VOL. 27, NO. 1, JANUARY 2012.
[2] Niv Goldenberg and Avishai Wool,”Accurate modeling of Mod-
bus/TCP for intrusion detection in SCADA systems”,international jour-
nal of critical infrastructure protection 6 (2013) 63–75.

You might also like