ACTOR CRITIC APPROACH BASED ANOMALY DETECTION FOR EDGE COMPUTING ENVIRONMENTS

International Journal of Computer Networks & Communications (IJCNC) Vol.15, No.1, January 2023
DOI: 10.5121/ijcnc.2023.15104 53
ACTOR CRITIC APPROACH BASED ANOMALY
DETECTION FOR EDGE
COMPUTING ENVIRONMENTS
Shruthi N1
and Siddesh G K2
1
Department of Electronics & Communication, BNMIT, Bangalore, Karnataka, INDIA
2
Research Head, New Horizon College of Engineering, Bangalore, Karnataka, INDIA.
ABSTRACT
The pivotal role of data security in mobile edge-computing environments forms the foundation for the
proposed work. Anomalies and outliers in the sensory data due to network attacks will be a prominent
concern in real time. Sensor samples will be considered from a set of sensors at a particular time instant as
far as the confidence level on the decision remains on par with the desired value. A “true” on the
hypothesis test eventually means that the sensor has shown signs of anomaly or abnormality and samples
have to be immediately ceased from being retrieved from the sensor. A deep learning Actor-Criticbased
Reinforcement algorithm proposed will be able to detect anomalies in the form of binary indicators and
hence decide when to withdraw from receiving further samples from specific sensors. The posterior trust
value influences the value of the confidence interval and hence the probability of anomaly detection. The
paper exercises a single-tailed normal function to determine the range of the posterior trust metric. The
decision taken by the prediction model will be able to detect anomalies with a good percentage of anomaly
detection accuracy.
KEYWORDS
Reinforcement Learning, Actor Critic, Security, Anomaly Detection, Posterior Belief
1. INTRODUCTION
Information and computing are ubiquitous in the sphere of communication. One may refer to
information being available and handled at the user devices while the other may refer to the
information at the core cloud infrastructure. Information also gushes its way through various
intermediate communication networks and servers. Challenges of IoT (Internet of Things) based
networks like real-time massive data generation, heterogeneous data, dynamic demeanor
networks, constrained memory, and resources are still difficult to elucidate. However, the
commendation goes to researchers who have shed light on how to effectively glean embedded
intelligence features from Machine Learning (ML), and Deep Learning (DL) techniques to
incorporate them in IoT devices and networks.
There will an explosive growth in computationally expensive sensor data due to the erratic rise in
sensory devices like wearable devices, smart phones, daily use appliances, vehicles, etc., Data
analysis and computing will be highly challenging. Captured data will most of the time be of high
dimension. Network handling capacity and handling delay mitigation requirements will pose
stringent challenges too. Real-time inferences and responses are critical in a majority of
applications. With more and more data being generated from the physical world, there is a
definite boost in bottlenecks concerning network bandwidth and transmission speed too.

54
The authors of [1] have consolidated the above challenges into mainly latency, scalability, and
privacy. Users send data to the cloud and this bears privacy concerns. Exposure to network
attacks is always at the higherside of probability when it comes to data handling. Anomaly
detection refers to the identification of an anomalous or abnormal activity in the EC network
predominantly due to an attack. Security and privacy assurance should be mitigated and handled
effectively in each every network zone/layer in all ”edge computing” system designs. The authors
of [2] clearly categorize the security challenges in the 4layers namely (a) core cloud (b) edge
servers (c) edge networks and (d) edge devices. The various attacks can include spoofing, Man-
in-the-Middle, Rogue, Denial of Service (DoS), Distributed Denial of Service (DDoS) and other
smart network attacks influencing the performance. However, our work concentrates on detection
ofanomalies and outliers with respect to attack detection irrespective of the type of attack. A
sustainable, rewarding edge computing model should abide by (a) Data Confidentiality, (b) Data
Integrity, (c) Authentication, (d) Access Control and (e) Privacy and Security.
Our work will mainly streamline the security of data based on anomaly detection. The work
explores the possible avenues of using cutting-edge deep learning algorithms in the security
domain, extensively for edge computing environments. Henceforth, the discussion will be
concerning the following four dimensions namely (i) Anomaly Detection (ii) Network Security
(iii) Edge Computing Environments (iv) Deep Learning based approaches and Algorithms.
1.1. Convergence of Edge Computing and Deep Learning
We now understand that a centralized, traditional cloud computing model is not very
encouraging. Edge Computing (EC) promises to be a viable alternative or can say a probable
solution to most of the network challenges. Edge Computing is a novel architecture wherein the
services of the cloud are protracted up to the network edge. This reduces latency. The platform is
maintained very close to the data source, this supports effective, speedy real-time processing,
data optimization, security, and privacy.[3]. The authors of [4] have highlighted the advantages of
edge computing as low latency, energy saving, context awareness, privacy, and security. Scale
reduction in edge computing makes these networks less prone to attacks as compared to the large-
scale data centers of the cloud.
The image in Fig. 1 represents the exposure of edge computing architecture to different network
attacks. Various distributed, dispersed, appropriated low latency and dependable astute services
alongside DL are ensured by the edge networks. The computational efficiency and edge user
experience for time-critical applications improve and localized user experience could be
improved significantly with edge computing when they discuss cloud offloading.
Deep Learning (DL) is a catalyst for edge intelligence DL is a machine learning technique that
has originated from Artificial Neural Networks (ANN).

55
Figure 1. Edge Computing Environment subject to attacks
The ”DL - edge computing” combo is guile but in-depth understanding and interpretation of DL
models, edge computing features and also design, development and deployment standards is a
major requirement. DL is known for distributed computing and, analysis of unlabelled,
uncategorized, and unsupervised data [5].There are many real-time complex machine learning
tasks which utilize deep learning have demonstrated good performance, heeding to privacy and
latency concerns.
The category of reinforcement learning (RL) algorithms works towards securing data transfer at
the edge environments. RL works on the principle of ”reward function” system. The agent learns
from the feedback post its interaction with the environment. Agent’s action will be evaluated
based on the reward. This kind of a DL technique is highly suggested for dynamic edge
computing environments wherein learnings are based on experiences rather than on the data.
A RL environment is usually cast as a Markovian Decision Process (MDP) which synthesizes the
entire concept based on a generic computational framework which will be used by the model to
draw conclusions or otherwise decisions.
The authors of [6] summarize few compelling RL based security methods which promise to
better the security parameter in the ”edge”. However, RL-based edge security solutions still
welcome a lot of challenges which are to be addressed.
Q-learning is a prominent RL technique which has proved its competency in spoofing, malware,
jamming and eavesdropping management. Q-values revolving around the learning rate and
discount factor are updated by the iterative Bellman equation. Learning performance witnesses an

56
uplift for stochastic environments. Unfortunately, edge security solutions implemented via Q-
learning face high-ambit issues. The ”model-less”, online learning method has no advance
information about the working environment unlike ”Markovian”. Q-learning variants like the
”Dyna Q” offer to be a savior then. Finally, the conclusion statement that can be put across is that
both MDP and Qlearning project higher culminations as against convex optimization. Also,
consummation of Q-learning will be unstable for lesser number of learning times. [7].
The authors of [8] mention about security defense mechanism in Deep Neural Network based
systems. However, the work tries to focus on providing an authentication strategy based on
Reinforcement learning. The proposed work focuses on the Actor Critic (AC) algorithm to detect
anomalies in the network. DNNs proximate the state-action duo’s value function. The critic
associated gradient information is available for analysis. It is observed that especially for real-
time random processes and continuous domain environments, model approaches based on
policies and protocols would be a good bet. Even though value-based approaches are stable and
sample efficient, the former will have a very impressive, faster convergence rate. As far as the
”state-action duo” values are continuous space entities in seemingly stationary environments. A
simple working idea of Actor-Critic is depicted in the Fig. 2 below.
Figure 2. Actor-Critic environment.
1.2. Network Architecture for MEC Security
A better option in terms of latency and security is to offload data at the resource efficient edge
servers which are well equipped to execute the DNN models that are provided. The deliverable
capabilities of deep learning algorithms can be experienced if the challenges with respect to edge
devices and the edge environment as a whole are made to move towards efficient solutions. The
authors of [9] provide an insight into few of the challenges in the era of edge computing. The
choice of a deep neural network, model partition and allocation scheme all have a role to play in
security and privacy preservation.
A specialized DNN suited for temporal data is the Recurrent Neural Network (RNN)which has
loops in their layer connections so as to store the state value and envision the sequential inputs
[10].
The prior computed values influence the output. The “memory” catches in RNNs resorts to
confiscate the information calculated so far. RNN architecture supports data persistence and
models short term dependencies in a fine-tuned manner. The shortcomings of RNN include the
hardship in training, the exploding-vanishing gradients during training, inability to handle very
long sequences in the case of tanh and relu used as activation functions. Long-term dependencies
will also impact good results.

57
The ”Vanishing Gradient” problem seeks a good solution from Long Short-Term Memory
(LSTM) networks. The gradient can pick up patterns over larger regions of the sequence. LSTM
networks have memory blocks that are connected into layers instead of neurons. The Forget,
Input and Output gates constitute the total working of the LSTM model. The Forget gate
discovers what details are to be discarded from the block. The Forget gate reviews at the previous
state and the input, delivers a value well lesser than 1 for each number in the cell state.
The input gate identifies which value from input should be used to modify the memory/ store in
the cell state. First, a sigmoid layer called the “input gate layer” decides which values to update.
Next, a tenth layer creates a vector of new candidate values, that could be added to the state. In
the next step, these two values will be combined to create an update to the state.
The role of the sigmoid layer is to check and deliver selected parts of the cell state to the output
end. The cell state is cascaded with ”tenth”, following which the gate output is multiplied with
the cascaded factor.
1.3. Contributions
The contribution of the proposed work is oriented towards providing a multiple layer-based
implementation of a security model. The criticalities of security in communication networks are
analysed via detection of anomalies with the aid of a powerful deep learning algorithm - The
Actor Critic algorithm.
1) The model uses of LSTM based Recurrent Neural Networks as against Convolutional
Neural Networks for active hypothesis testing to capture the temporal dynamics of the
sensory environment. The attack detection model is a binary indicator, based on which the
system decides when to stop receiving sensor data from a specific sensor.
2) Actor Critic algorithm is implemented with both actor and critic taken as separate networks
which interact with each other to improvise on the detection accuracy of network anomalies.
The work banks upon the” criticise-to-improve” working principle of the algorithm.
1.4. Organization of the Paper
Section III briefly explains the proposed system model which includes the problem formulation,
and establishment of the hypothesis. The section also discusses how to solve the active
hypothesis testing problem and the selection of posterior belief thresholds. Section IV reveals the
deep actor-critic framework for optimal policy selection.MDP Parameterization, modeling
equations of both Actor and Critic framework, and reward computation are highlighted in the
section. Section V gives an outline of the algorithm. Details of the confusion matrix are explained
in Section VI. Section VII deals with the results and evaluation. The conclusion and opportunity
for future work are enclosed in Section VIII.
2. METHODOLOGY
2.1. Anomaly Detection Environment
The correlation between features is one of the main reasons for choosing the LSTM neural
network. Recurrent neural networks are a wise option to handle use cases of hypothesis testing
which will be discussed elaborately in the upcoming sections. Traditional time-series models
analyze time-series as separate entities and do not consider the complex inter dependencies
among the different timeseries [11]. Also, the Q-function is approximated by LSTM in order to

58
tackle the attacking agents due to its potential to capture the temporal dynamics of the
environment [12].The proposed model will determine whether an anomaly has been detected in
the input. Eventually, output of the attack detection model should be a binary value that indicates
’1’ in case of attack and ’0’ in case of no attack. Input database can be split as training and testing
data sets. The core architecture includes the three vital layers of LSTM. A suitable activation
function and loss function is chosen in the FC layer. Dropout will be used to mitigate overfitting
issues of the anomaly model. The model can be improved using an apt optimizer as well. The
closure of the model is taken care by a single neuron output layer which indicates the status of
attack. The Fig. 3 below represents the feasible LSTM architecture for anomaly or attack
detection.
Fig. 3. LSTM architecture for anomaly or attack detection.
The proposed work considers the fact that edge data will be fetched from various sources through
sensors in real-time. Sensors are supposed to provide impeccable data to the receiver end.
Sensors are many a time less reliable in content delivery either due to sensor construction,
environmental changes, lack of sensor calibration. However, our discussion is confined to non-
reliability of sensor data due to network attacks only.
The model proposed aims at a trust/ satisfaction metric which will vary proportionally with the
user provided QoS which indirectly is a measure of successful detection accuracy. A potent
reinforcement algorithm, the Actor-Critic is used to dynamically help the agent stretch out
towards a maximized reward. The agent’s main target will be to maximize the average value of
the user response in terms of the satisfaction metric (=), which is part of the reward calculation.
Each time the agent performs an action on the environment, the critic evaluates/criticises the
actor’s performance. The fundamental goals of a lucrative agent can be the following: (1)
Maximize the average reward function/ trust metric (2) Optimizing latency with respect to

59
offloading in MEC environments (3) Reducing the stopping time for a quick turnaround on the
sensed data [13]. However, the work focuses on only maximizing the trust metric so as to
bring in the security feature into the model.
Considering each independent sensor source as an independent process, we will look forward to a
model which can handle (P) processes. Inherently, the word ”process” refers to the ”sensor”. It is
obvious that each process is random and the probability of (say p) processes out of these (N)
processes being anomalous/ atypical (ρad) is acceptable. At a given time instant (t), (N) can take
values in the range [0,P].The assumption is that (N) does not exceed the value of (P) at any point
in time. The decision taken by the prediction model must be able to detect the anomaly with an
appreciable percentage of anomaly detection accuracy.
There will be 2P values that can fit into the state vector say (V) for (P) processes/ sensors. The
paper brings in the concept of state and action since they are required to be used in the Actor-
Critic implementation. The vector states can be represented by sample space V .The
algorithm will cater to all available sensors hence facilitating 2P −1 actions [14].
2.2. Problem Formulation– The System Model
Trust metric and confidence interval/satisfaction metric are the deciding factors for maximization
of reward in the proposed work. The satisfaction metric is the Bayesian log-likelihood ratio of
hypothesis at time (t) given as:
𝕴q (𝜷) = log
𝜷(𝒒)
𝟏−𝜷(𝒒)
(1)
Here, βq is the qth entry of a posterior belief vector β.
The average Bayesian log likelihood ratio contributes to the reward component. Rewards can be
averaged as:
𝑹𝝁(𝒕) ∶ =
𝟏
𝝉
∑ 𝑬𝝁
[ 𝒓𝝁 (𝒕)]
𝝉 −𝟏
𝒕 = 𝟏 (2)
where rµ(t) is the instantaneous reward of MDP given by:
𝒓𝝁 (𝒕) ∶ = 𝕴𝒂𝒗𝒈 (𝜷(𝒕)) − 𝕴𝒂𝒗𝒈 (𝜷(𝒕 − 𝟏)) (3)
Posterior belief threshold influences the probability of anomalous behaviour which is defined
with the help of a single-tailed function. The hypothesis tests true when anomaly is detected and
anomaly is detected when posterior trust value is within the range say (α,βmax) which define the
lower and upper bound of posterior values.
𝝆𝒂𝒅 (𝜷) = {
𝑭(𝜶) 𝜷 < 𝛼 − ℎ𝑦𝑝𝑜𝑡ℎ𝑒𝑠𝑖𝑠 𝑓𝑎𝑖𝑙𝑢𝑟𝑒
𝑭(𝜷𝒎𝒂𝒙) 𝜶 ≤ 𝜷 < 𝜷𝒎𝒂𝒙 − 𝒉𝒚𝒑𝒐𝒕𝒉𝒆𝒔𝒊𝒔 𝒔𝒖𝒄𝒄𝒆𝒔𝒔
(4)
The aim of the work in terms of the algorithm is to maximize the instantaneous reward of the
MDP rµ(t) depending on the average Bayesian log likelihood ratio =avg(β). The algorithm ensures
to minimize the squared value of temporal error ψµφ(t). The Actor update towards the target
value follows.

60
2.3. Establishment of Hypothesis
Sensor data can be used to assess the believability and validity of a hypothesis. This is what is
referred to as ”Hypothesis testing.” Herman Chernoff was one of the first to formulate active
hypothesis testing [15]. The model has to administer to infer the hypothesis for a given set of
observations and learn the optimal selection policy. At time (t), we define a sample space ζ (t)
which shall constitute all samples that are captured from different sensors at that time instant
(t).Let us denote the set by say {O1,O2,.....OP}.Therefore, the corresponding observations of each
sensor at time (t) can be represented as O where a(t) refers to the process observed at
each sensor given as a(t) 1,2,3,....P}.We can refer to the sample from a particular sensor or a
particular process (i) as well.
The active hypothesis testing problem equivalent to the anomaly detection problem has 2P
hypothesis. Out of all the unknown number of processes, a hypothesis test will be defined as
Hqfor all values varying from q = 1,2,.....2P. It is to be noted that out of all samples on the
observation space, samples will be collected from the sensor as long as hypothesis Hqis verified
to be logically true and the others when the hypothesis Hqis false. If at any time the hypothesis
tests “true”, that eventually means that the process has tested positive for anomaly or abnormality
and samples have to be immediately stopped from being retrieved from sensor. The supply/sensor
chain has to be terminated.
2.4. Establishment Solving the Active Hypothesis Testing Problem
Each of the possible states of the processes are associated with a hypothesis which is proposed by
the decision maker. Later, the posterior probabilities are computed based on the hypothesis laid
out. Using all the information, the agent forms a posterior trust value β(t) on hypothesis Hqat time
(t). The observation information available with the agent at any time (t) is given by:
𝑶𝝉 ⊆ [𝑷]𝟏: 𝒕−𝟏 (5)
Sequence of actions selected by the agent based on the critic’s feedback can be represented as:
𝑨𝝉 ⊆ [𝑷]𝟏: 𝒕−𝟏 (6)
The optimal policy design depends upon the estimation algorithm. One can foresee two different
aspects of the belief vector. In the first case, the belief vector is the posterior probability that the
jthprocess is non-anomalous. Therefore, the trust vector can be updated on the arrival of each
observation as the probability of the state being ’0’, as shown in the equation.
𝜷𝒋(𝒕) = 𝝆(𝒔𝒋 = 𝟎 | 𝑶𝝉,𝑨𝝉; 𝝉 = 𝟏, 𝟐,𝟑, . . . (𝒕 − 𝟏)) (7)
In the second case, we denote βq(t) as the posterior trust value of the hypothesis Hqbeing true at
time (t) such that βq(t) ∈ [0,1]2P.
𝜷𝒒(𝒕) = 𝝆(𝑯 = 𝒒 | 𝑶𝝉,𝑨𝝉; 𝝉 = 𝟏,𝟐, 𝟑,. . . (𝒕 − 𝟏)) (8)
Bayes rule administers a way to positively refurbish the beliefs based on the arrival of fresh
samples of evidence. In our case, we were trying to compute the probability value if and only if
the given hypothesis is considered true. Given additional evidence such as the sequence of
actions, we can update our probability. Estimations can be improved based on the effective usage
of the available prior knowledge. The general final form of Bayes rule is formulated as:

61
𝝆 (𝑨|𝑩) =
𝝆(𝑨).𝝆(𝑩|𝑨)
∑ 𝝆(𝑨𝒒).𝝆(𝑩|𝑨𝒒)
𝑯
𝒒=𝟏
(9)
The final form of Bayes rule can be applied to our study by replacing (A) with hypothesis and (B)
with the corresponding observations.
𝜷𝒒(𝒕) =
𝝆(𝑯=𝒒).𝝆(𝒁[𝑨 𝝉(𝝉)]|(𝑯=𝒒)
∑ (𝑯=𝒒).𝝆(𝒁[𝑨𝝉 (𝝉)]|(𝑯=𝒒)
𝑯
𝒒=𝟏
(10)
It is to be noted that τ varies up to the value of (t-1) starting from unity and Z[Aτ(τ)] is the
corresponding sensor measurement or the corresponding observation. Observation probabilities
corresponding to one of the hypotheses are segregated by the decision maker, taking into account
that the value exceeds the desired confidence level βhigh.
Confidence intervals are widely accepted as a preferred way to present study results. Confidence
intervals measure the uncertainty of an estimate. Maximum Likelihood Estimation (MLE)
methods provide 2 major approaches to construct the confidence interval. The first being the
”Asymptotic” method wherein the likelihood function can be approximated by a parabola around
the estimated parameters. Assumption is its validity for very large samples; otherwise, the
performance of the confidence interval will be poor, considerably less than the nominal rate of
95%. Confidence limits (βlow, βhigh) reflecting the asymmetric sampling distribution of data may
be based on different variance estimates [16]. The second approach is the Likelihood Ratio
Test(LRT), an experiment performed by R.A.Fisher to find the best overall parameters value and
likelihood. LRT is range preserving and behaves in a predictable manner as the sample size
grows.
The confidence interval has two unusual features: (i) The endpoints can stray outside the
parameter space; that is, one can either get βlow <0 or βhigh >1. (ii) The confidence interval width
becomes zero if no successes or failures are observed.
A possible solution to ensure that the endpoints do not cross over the parameter space is to
parameterize using a completely different scale like”log-odds/ logit/ logistic transformation”. The
logit function is the inverse of the logistic sigmoid function 1/(1+e−x). It is extensively used in
statistics and machine learning to quantify confidence level [17].
The intent is to evolve with a compelling selection strategy or an optimal policy which the
actor/agent can implement and strive towards an increased value of confidence level = on the
hypothesis (H) being true. In response to all the sensor measurements, one can perform the
reward calculation. We define the average Bayesian log likelihood ratio as:
𝕴𝒂𝒗𝒈(𝜷) = ∑ 𝕴𝒒(𝜷). 𝜷𝒒
𝑯
𝒒=𝟏 (11)
Asymptotic expected reward is based on the average rate of increase in the confidence level on
the true hypothesis H and is defined as [18]:
R(𝝁) ∶= 𝐥𝐢𝐦
𝑶𝝉→∞
𝟏
𝑶𝝉
𝑬𝝁
[𝕴(𝜷(𝑶𝝉 + 𝟏) − 𝕴(𝜷(𝟏)] (12)
However, the proposed work extends the concept by averaging over all the measurements
obtained at different time intervals for a particular sensor. Maximizing the reward can be devised
as an infinite horizon, average-cost Markovian Decision Process (MDP) problem. The posterior
trust vector β(t) will act as the state of the MDP. Henceforth, the average of rewards can be
formulated as in (2).

62
2.5. Selection of Posterior Belief Thresholds
It has been highlighted earlier those sensory measurements from various sensors are stochastic
and random in nature. Henceforth, it would be highly likely to define the probability (ρad) of
anomalous/ atypical behavior in the data indicating the influence of a network attack.
For an anomaly detection task, the posterior trust value (β) influences the value of the confidence
interval and hence even the probability of attack/ anomaly detection. As (β) increases, (ρad) also
escalates. According to the above analysis, we employ a singletailed normal function, ρad(β), to
reflect the relationship between the vehicle speed and the task delay constraint. The properties of
the one-tailed normal function are shown in Fig.4.
Fig. 4. Posterior Trust and Anomaly probability modeling for Anomaly detection.
When the posterior trust value is within the range of (0,α), the function ρad(β) is well within the
boundary of the hypothesis being rejected i.e; anomaly has not been detected. On the contrary,
When the posterior trust value is within the range of (α,βmax), the hypothesis is in acceptable
condition indicating that an anomaly has been detected. The algorithm has to now reach the
”termination” condition. It is suitable to use the following function to describe the Posterior Trust
and Anomaly probability model for anomaly detection as follows:
𝝆𝒂𝒅 (𝒋, 𝒕) = 𝒆𝒙𝒑 (
𝜷𝒋,𝒕
𝟐
− 𝜷𝒎𝒂𝒙
𝟐
𝟐𝜶𝟐 ) (13)
βmax ≈ 1 and in order to ensure that the probability that anomaly detection probability is within the
maximum value exceeds 95%, we denote the following [19]:
𝜶 =
𝜷𝒎𝒂𝒙
𝟏.𝟗𝟔
(14)
3. DEEP ACTOR-CRITIC FRAMEWORK FOR OPTIMAL POLICY SELECTION
This section describes the learning framework to solve the anomaly detection problem. The frame
of reference will bank on the advantages of a constructive-feedback oriented feedback critic
network, an advantage function and TD learning-based update. All these highlighted features are
promoted by the Actor-Critic framework wherein a trained critic model approximates the value
function. The ”advantage” function conveys to the agent about the quantum improvement

63
required in comparison with the average action value taken at that state. Alternately, this function
computes the extra reward that the agent gets if taking that action.
The Actor and Critic networks are discussed in detail in the subsections. The deep reinforcement
learning based algorithm will emphasize on how to interact with the environment effectively and
how to learn effectively from experiences.
3.1. MDP Parameterization
The underlying fundamental principle of the Actor-Critic is the MDP which is characterized by
four fundamental segments namely:
1) Time-indexed STATE variables St in finite state space (S) The posterior trust vector βt will be
the state of the MDP at time (t). We consider the time index since the β value changes
dynamically with each sensor measurement arriving at different time instants.
2) Time-indexed DECISION or ACTION variables at in finite action space (A) - The agent takes
the decision, an action (a) with the help of a constructive critic and transitions to a new state
β(s,a).
3) The optimal policy in general has to maximize the average REWARD value. Reward
calculation helps to evaluate the agent’s action on the environment. Specifically, in the work
considered, policy has to augment and escalate the averaged value of confidence level
metric =. This ensures that the decision accuracy is definitely boosted. maximize the average
confidence level metric = and hence maximize the decision accuracy.
3.2. Actor Framework
The Actor and Critic networks will be two separate neural networks which interact with each
other productively. The actor explores and learns the policy (µ) on the true hypothesis
characterized by say (φ) which maps the posterior trust distribution (β)to the action space i.e;
chooses a valid action based on the posterior probabilities (β).Valid action refers to selecting the
corresponding sensor and receiving the sample to update the posterior belief. In each iteration,
the agent will score all valid actions, and choose the one with the highest score to execute.
The actor network’s output layer provides an assemblage of all actions that belong to the action
space; the elements in this action space can be treated as the subset of [P].
𝒂𝒕 ∈ 𝑨 ∶= 𝝁𝝋 (𝜷(𝒕 − 𝟏)) (15)
3.3. Critic Framework
The critic portrays its network as a ”state-value function approximator”. In fact, the actor updates
its policy based on this value function computed by the critic.
The new state is subject critic-based computation and this happens after each action selection.
Results obtained can either be better than the prior or even worse. Critic receives the following
data as input:
1) Posterior trust value at time (t) - β(t)
2) Instantaneous Reward value rµ(t) from equation [11]
3) Posterior trust value at time (t − 1) - β(t − 1)

64
TD error can be used to evaluate the action implemented (at) for the particular state (St). If the
TD error is positive, it suggests that the inclination to (at) must be intensified for the upcoming
observations. On the contrary, for negative TD error, inclination must be weakened for future
values.
3.4. Bringing in Reward Computation
Critic is responsible for temporal error calculation and can be expressed mathematically as:
𝝍𝝁𝝋(𝒕) = 𝒓𝝁(𝒕) + 𝝀. 𝑽(𝜷(𝒕)) − 𝑽(𝜷(𝒕 − 𝟏)) (16)
In the above equation, (V) is the estimate of the current value function and (λ) is the discount
factor (0,1). Discount factors become important to be introduced especially in case of time-
invariant samples or stationary values. It is required to obtain an optimal value function to
determine the best fit line to data by using Mean Squared error (MSE). The critic updates itself
by minimizing the squared value of temporal error (ψ2).
3.5. Updating the Actor and it’s Parameters
Policy gradient helps the actor to get updated. TD error (ψ) from equation [16] will be used to
compute the gradient as depicted in the equation below:
𝛁𝝋. 𝑱(𝝋) = 𝑬𝝁𝝋 [𝛁𝝋. 𝒍𝒐𝒈 𝝁𝝋(𝜷(𝒕 − 𝟏),𝒂𝒕). 𝝍𝝁𝝋
] (17)
In the process of TD learning, the weights will get updated at each learning step in real-time. The
technique does not wait till the end of the episode to complete update of the weights. This
supreme characteristic overcomes the shortcomings of the policy gradients. µφ(β(t − 1),at) refers
to the score for the selected policy i.e; how well did the actor faired with the current policy, it’s a
sort of evaluation. ∇φ represents the gradient with respect to the parameter φ.
The actor also is responsible for updating φ by taking the support of the policy gradient and TD
error value.
𝝋𝒕 = 𝝋𝒕−𝟏 + 𝝍𝒕−𝟏. 𝛁𝝋𝒕−𝟏
[𝒍𝒐𝒈 𝑽𝝋(𝑨𝒕| 𝜷(𝒕 − 𝟏))] (18)
Now, the paper denotes the weighted difference of actor parameters where η is the learning rate
with values in the range set (0,1).
𝜼. [𝛁𝝋. 𝒍𝒐𝒈 𝝁𝝋 ( 𝜷(𝒕 − 𝟏),𝒂𝒕). 𝝍𝝁𝝋
] (19)
Actor can now accept the criticism (positive/ negative) which has been provided by the critic
network, update itself as below and work towards the target value to reach the optimal goal.
𝝋𝒕 = 𝝋𝒕−𝟏 + 𝜼.[𝛁𝝋. 𝒍𝒐𝒈𝝁𝝋(𝜷(𝒕 − 𝟏), 𝒂𝒕).𝝍𝝁𝝋] (20)
The dual networks will be handled and trained solely by disembarking on the gradient value. This
definitely ensures traversing in the direction in which value function is accelerated [20] (to find
the global maximum) to update both their weights.

65
4. ALGORITHM OVER VIEW
5. INTERPRETATION OF THE CONFUSION MATRIX
The confusion matrix can be used effectively in machine learning to visualize vital predictive
analytical parameters like accuracy, F1 score, recall score and precision score. Analysis of True
Negatives (TN), True Positives (TP), False Positives (FP) and False Negatives (FN) help to
measure the performance of the anomaly detection system.
Accuracy =
𝑻𝑷+𝑻𝑵
𝑻𝑷+𝑭𝑷+𝑻𝑵+𝑭𝑵
Precision =
𝑻𝑷
𝑻𝑷+𝑭𝑷
Recall =
𝑻𝑷
𝑻𝑷+𝑭𝑵
F1 =
𝑷𝒓𝒆𝒄𝒊𝒔𝒊𝒐𝒏 ∗ 𝑹𝒆𝒄𝒂𝒍𝒍 ∗ 𝟐
𝑷𝒓𝒆𝒄𝒊𝒔𝒊𝒐𝒏 + 𝑹𝒆𝒄𝒂𝒍𝒍

66
6. RESULTS AND DISCUSSION
6.1. Selection of Dataset
Malicious input detection systems are trained on internet traffic record data. This research work
makes use of NSL-KDD which is the most prevalent data set. The work has taken nearly
1,40,000 sample points maintaining a workable ratio of training and testing data as approximately
82% and 18% respectively. The subset datasets used are KDDTrain+ and KDDTest+ .The dataset
samples have a combination of normal, Denial-of-service, Probe, User-to-Root and Remote-to-
Local attacks. The data processing logic obtains n-row batch from the dataset and returns
dataframe with correct labels for detection.
6.2. Selection of Hyperparameters
RMS Prop optimizer is used to accelerate the optimization process. The proposed work uses
ReLu as the activation function. Table II shows a list of hyperparameters used and their
respective values.

67
6.3. Simulation Results
Fig.5 represents the average value of accuracy of detection of network attacks. The bar plot gives
a comparative visualization of correct estimates, false positives and false negatives. False
negatives refer to those sensor samples which are attacked and yet are left undetected; this is a
serious threat to the system performance. However, a marginally better detection accuracy is
witnessed for the normal, DoS and Probe attacks.
Table I : hyperparameters with values
Hyperparameters Values
Batch size 100
Activation function ReLu
Optimizer RMSProp
Loss Function MSE
Discount factor 0.99
Learning Rate 0.00025
Fig. 5. Detection accuracy of Network Attacks in NSL-KDD
Fig.6 depicts the colour encoded heat map to represent the confusion matrix. Best True Positive
value achieved for normal attacks according to the chart is nearly 0.96 .
Fig. 6. Confusion Matrix of True versus Predicted values of attack detection.

68
Fig.7 depicts the colour encoded heat map to represent the action probabilities with respect to the
states. The policy prediction function uses the Bayesian probability values and computes the
action probabilities.
Fig.8 depicts the graph to represent the Bayesian Posterior probabilities which have been
computed using the relation mentioned below.
P(𝝋|𝒚) = 𝒅𝒊𝒔𝒕. 𝒑𝒅𝒇(𝒙,𝒂𝒑𝒓𝒊𝒐𝒓 + 𝒅𝒂𝒕𝒂,𝒃𝒑𝒓𝒊𝒐𝒓 + N – data) (21)
The Bayes theorem is used, priori probability is updated and then the value of posteriori
probability is obtained. PDF of the posterior distribution must be the Probability Density
Function (PDF) of the Beta distribution. Proportionality constant must be whatever constant is
required to make this PDF integrate to 1 over p (0, 1).
Fig. 7. Variation of action probabilities.
Fig. 8. Variation of Bayesian probabilities.
Fig.9 depicts the representation which demonstrates the variation of rewards and accuracy with
respect to varying episode lengths. Rewards and accuracy values are highlighted separately in the
graph. Bayes theorem will be used to establish independence between the system inputs and
model parameters.

69
Fig. 9. Reward Analysis with respect to Episode Length
6.4. Discussion of Results
1) The analysis of Correct Estimates with respect to false negatives and false positives
has shown a successful detection accuracy in case of Normal, DoS and Probe
attacks.
2) Metrics of TP, FN and FP have been obtained and tabulated below for Normal, DoS
and Probe attacks.
TABLE II VALUES OF VITAL METRICS
Name of Metric TP FN FP TN
Normal 0.96 0.04 1.92 2.17
DoS `0.88 0.12 0.29 3.83
Probe 0.72 0.28 0.05 3.1
7. CONCLUSION AND FUTURE WORK
The proposed work implements the Actor Critic algorithm for anomaly detection. The deep
reinforcement learning approach incorporates a LSTM network based Actor Critic algorithm for
efficient anomaly detection. The algorithm presented aims to identify and fix the incongruous
processes among a collection of binary processes. The analysis picks a single process at a time.
The samples from sensors are further ceased from being received once anomaly has been
detected.
The satisfaction metric which is the Bayesian log likelihood ratio of hypothesis is the determining
factor for reward maximization. The hypothesis tests true on anomalous detection when posterior
trust value is within the range defined by the bounding posterior values. Computations are based
on marginal probabilities of the processes. The one-tailed normal function used in the proposed
model helps to decide the range of the posterior trust metric. The range influences the decision of
the hypothesis being rejected or accepted.

70
The scope for future work aims at minimizing the values of false negatives so as accelerate the
detection accuracy. An optimal process selection policy can be proposed with an improvised deep
reinforcement learning based Asynchronous Advantage Actor Critic (A3C) algorithm. The
algorithm can be implemented to predict both value function and optimal policy functions as is
done here using Actor Critic.
ACKNOWLEDGMENT
This research paper would not have been possible without the exceptional support of my mentor,
Dr. Anitha V who is a wonderful inspiration. Her enthusiasm, knowledge share, domain expertise,
impeccable inputs and willingness to give her time have been valuable. The authors would like to
thank her for all her contributions.
BIOGRAPHIES
Shruthi N is a Ph.D. research scholar in Bangalore, Karnataka, India under Visvesvaraya
Technological University, Karnataka, India. She received a Bachelor’s degree in Electronics
Communication Engineering and a Master’s degree in Digital Electronics Communication in
2005 and 2014 respectively. Her areas of interest are Network Security, IoT and Embedded
Systems. She has nearly 4 years of industry experience and 8.5 years of teaching experience with
6 International Journal publications to her credit.
Dr.Siddesh.G.K. is Head, Research and Development at New Horizon College of Engineering,
Bangalore. He received a Bachelor’s degree in Electronics Communication Engineering from
Bangalore University in 1998, M.Tech. in Digital Electronics and Advanced Communications
from Manipal Institute of Technology, Manipal, Karnataka in 2002 and Ph.D.in Electronics
Communication Engineering from Visvesvaraya Technological University, Belagavi in 2013.
His work experience includes academic, research administration of more than 20+ years in
various engineering colleges. He has published more than 45 research papers in various National,
International Journals and Conferences in India and abroad. He also has book chapters from
reputed publishers to his credit.
DECLARATION
Compliance with Ethical Standards: The authors declare that the submitted manuscript has not
been published elsewhere. No funding was received to assist with the preparation of this
manuscript. All authors certify that they have no affiliations with or involvement in any
organization or entity with any financial interest or non-financial interest in the subject matter or
materials discussed in this manuscript. Competing Interests: The authors have no competing
interests to declare that are relevant to the content of this article.
Research Data Policy Data Availability: The datasets generated during and/or analysed during
the current study are available from the corresponding author on reasonable request.
Conflicts of Interest: The authors have no competing interests to declare that are relevant to the
content of this article. There is no conflict of interest.

71
REFERENCES
[1] J. Chen and X. Ran, “Deep learning with edge computing: A review.,” Proc. IEEE, vol. 107, no. 8,
pp. 1655–1674, 2019.
[2] J. Zhang, B. Chen, Y. Zhao, X. Cheng, and F. Hu, “Data security and privacy-preserving in edge
computing paradigm: Survey and open issues,” IEEE access, vol. 6, pp. 18209–18237, 2018.
[3] S. Wang, X. Zhang, Y. Zhang, L. Wang, J. Yang, and W. Wang, “A survey on mobile edge networks:
Convergence of computing, caching and communications,” Ieee Access, vol. 5, pp. 6757–6779, 2017.
[4] Fangxin Wang, Miao Zhang, X.Wang, X.Ma, J.Liu, ”Deep Learning for Edge Computing
Applications: A State-of-the-Art Survey, “IEEE access, vol. 8, pp. 58322-58336, March 23 2020.
[5] F. Hussain, R. Hussain, S. A. Hassan, and E. Hossain, “Machine learning in iot security: Current
solutions and future challenges,” IEEE Communications Surveys & Tutorials, vol. 22, no. 3, pp.
1686–1721, 2020.
[6] L. Xiao, X. Wan, C. Dai, X. Du, X. Chen, and M. Guizani, “Security in mobile edge caching with
reinforcement learning,” IEEE Wireless Communications, vol. 25, no. 3, pp. 116–122, 2018.
[7] B. Cao, L. Zhang, Y. Li, D. Feng, and W. Cao, “Intelligent offloading in multi-access edge
computing: A state-of-the-art review and framework,” IEEE Communications Magazine, vol. 57, no.
3, pp. 56–62, 2019.
[8] A. Marchisio, M. A. Hanif, F. Khalid, G. Plastiras, C. Kyrkou,T.Theocharides, and M.
Shafique,“Deep learning for edge computing:Current trends, cross-layer optimizations, and open
research challenges,” Research Gate, IEEE, 2019.
[9] M. Zhang, F. Zhang, N. D. Lane, Y. Shu, X. Zeng, B. Fang, S. Yan, and
H. Xu, “Deep learning in the era of edge computing: Challenges and
opportunities,” Fog Computing: Theory and Practice, pp. 67–78, 2020.
[10] I. Goodfellow, Y. Bengio, and A. Courville, Deep learning. MIT press, 2016.
[11] Z. Khan, M. Chowdhury, M. Islam, C.-Y. Huang, and M. Rahman, “Long short-term memory neural
networks for false information attack detection in software-defined in-vehicle network,” arXiv
preprint arXiv:1906.10203, 2019.
[12] T. T. Nguyen and V. J. Reddi, “Deep reinforcement learning for cyber security,” arXiv preprint
arXiv:1906.05799, 2019.
[13] C. Zhong, M. C. Gursoy, and S. Velipasalar, “Deep actor-critic reinforcement learning for anomaly
detection,” in 2019 IEEE Global Communications Conference (GLOBECOM), pp. 1–6, IEEE, 2019.
[14] G. Joseph, M. C. Gursoy, and P. K. Varshney, “Anomaly detection under controlled sensing using
actor-critic reinforcement learning,” in 2020 IEEE 21st International Workshop on Signal Processing
Advances in Wireless Communications (SPAWC), pp. 1–5, IEEE, 2020.
[15] H. Chernoff, “Sequential design of experiments,” The Annals of Mathematical Statistics, vol. 30, no.
3, pp. 755–770, 1959.
[16] G. Y. Zou, “Toward using confidence intervals to compare correlations.,” Psychological methods,
vol. 12, no. 4, p. 399, 2007.
[17] D. W. Hosmer Jr, S. Lemeshow, and R. X. Sturdivant, Applied logistic regression, vol. 398. John
Wiley & Sons, 2013.
[18] D. Kartik, E. Sabir, U. Mitra, and P. Natarajan, “Policy design for active sequential hypothesis testing
using deep learning,” in 2018 56th Annual Allerton Conference on Communication, Control, and
Computing (Allerton), pp. 741–748, IEEE, 2018.
[19] X. Huang, L. He, X. Chen, L. Wang, and F. Li, “Revenue and energy efficiency-driven delay
constrained computing task offloading and resource allocation in a vehicular edge computing
network: A deep reinforcement learning approach,” arXiv preprint arXiv:2010.08119, 2020.
[20] G. Joseph, M. C. Gursoy, and P. K. Varshney, “A scalable algorithm for anomaly detection via
learning-based controlled sensing,” arXiv preprint arXiv:2105.06289, 2021.

ACTOR CRITIC APPROACH BASED ANOMALY DETECTION FOR EDGE COMPUTING ENVIRONMENTS

More Related Content

Similar to ACTOR CRITIC APPROACH BASED ANOMALY DETECTION FOR EDGE COMPUTING ENVIRONMENTS (20)

More from IJCNCJournal (20)

Recently uploaded (20)

ACTOR CRITIC APPROACH BASED ANOMALY DETECTION FOR EDGE COMPUTING ENVIRONMENTS