SIOPA-DLMUC: A Self-Improved Orca Predation Algorithm with Deep Learning for Enhancing 5G Enabled Cognitive Radio Network Security

1.jpg

In the quest for more robust spectrum sensing capabilities, central and distributed collaborative networks have emerged as promising approaches [6]. In the central network, secondary users (SUs) work together by sharing their sensing information with a Fusion Center (FC) that aggregates sensing reports from all SUs to make an optimum decision regarding spectrum occupancy of primary users (PUs) [7]. Conversely, in the distributed network, SUs collaborates by sharing their sensing data amongst themselves, independently making final decision on PU spectrum occupancy without need for FC communication [8]. While collaborative networks offer significant advantages, they are liable to possible threats posed by malicious users (MUs), who may engage in unwanted intrusions among PUs and SUs, thereby compromising accurateness of spectrum sensing process.

To address these challenges, recent advancements have seen the utilization of Machine Learning (ML) methods, which aim to mitigate some of the limitations associated with traditional approaches [9]. ML, a prominent area within Artificial Intelligence, empowers machines with ability to learn autonomously. ML models are supervised learning, which involves a training procedure with labeled input data, and unsupervised learning, where training occurs with unlabeled input data [10]. These ML methods offer promising avenues for enhancing spectrum sensing in collaborative networks, thereby improving the overall resilience and accuracy of the system.

In this study, we develop a new Self‐Improved Orca Predation Algorithm with Deep Learning-Driven Malicious User Detection (SIOPA-DLMUC) technique for 5G assisted CRNs. The main goal of SIOPA-DLMUC system is to classify and detect occurrence of MUs in the CRN. The SIOPA-DLMUC technique involves a long short-term memory (LSTM) model for detection of MUs in CRN. In addition, SIOPA-based hyperparameter tuning process can be executed to improve performance of LSTM approach. To observe outcome of SIOPA-DLMUC system, a comprehensive set of simulations is carried out on our database, comprising four kinds of attacks namely byzantine attack, jamming attack, SSDF attack, and Primary User Emulation (PUE) attack with normal samples.

In this article, we introduced an automated solution, the Self‐Improved Orca Predation Algorithm with Deep Learning Driven Malicious User Detection (SIOPA-DLMUC), tailored to address the pressing challenge of detecting Malicious Users (MUs) within the context of 5G-assisted Cognitive Radio Networks (CRN) susceptible to various attacks. The SIOPA-DLMUC operates through two critical stages: firstly, it harnesses Long Short-Term Memory (LSTM) to detect MUs effectively, offering the capability to identify temporal patterns associated with Byzantine attacks, Jamming Attacks, Spectrum Sensing Data Falsification (SSDF) attacks, and Primary User Emulation (PUE) attacks, as well as normal network behavior. The second stage utilizes the SIOPA to fine-tune hyperparameters, enhancing the model's adaptability to detect these specific attacks. SIOPA leverages advanced features, including an acceleration mechanism, memory, learning, and social interactions, to navigate the complex landscape of CRNs under the influence of these disruptive attacks. The model also incorporates a 'BubbleNet' mechanism that refines encircling behaviors, bolstering its ability to detect MUs and ensuring the security of CRNs operating within the dynamic 5G environment.

The SIOPA-DLMUC system aims to classify and detect the occurrence of the MUs in the CRN. To achieve this, SIOPA-DLMUC algorithm follows two phases of processes namely LSTM-based detection and SIOPA-based hyperparameter tuning. Figure 2 portrays workflow of SIOPA-DLMUC methodology.

3.1 Stage I: MU detection process

The detection of the MUs in the CRN is performed by the use of the LSTM model. LSTM works by employing a gating mechanism, consisting of forget, input, and output gates, to selectively process and store information in memory cells, allowing it to effectively capture and analyze temporal patterns in the data [19].

The LSTM is a kind of in-between NN model. LSTM comprises 4 neural system layers that interface in an optimum manner. LSTM adds or removes information to memory unit, using the "Gateway". It includes multiplication and layering work of sigmoid NN. The sigmoid layer opposites feature data by sigmoid ability and estimates outcome in range of [0, 1], representing what amount of information elements can be experienced. “1” represents that each data is permitted that sent. “0” denotes that no date is allowed that sent. The gating mechanism in the LSTM is related to the forgetting gate information gateway and an output gate.

The Forget gate selects what data to be discarded or retained from the memory:

$F_G=\sigma\left[w^F\left(F_t, Y_{t-1}\right)+c^F\right]$              (1)

In Eq. (1), $F_G$ is forget gate. C and w are controlled and weighted borders. $F_t$ addresses input at existing timestamp; $Y_t-1$ shows outcome at $t-1$ timestamp in prior square of LSTM. $\sigma$ shows sigmoid function.

The input gate $I_G$ selects the data that must be saved:

$I_G=\sigma\left[w^I\left(F_t, Y_{t-1}\right)+C^I\right]$             (2)

Lastly, the resultant gate defines which portion of the memory can be retained:

$O_G=\sigma\left[w^o\left(F_t, Y_{t-1}\right)+c^o\right]$            (3)

Additional candidate memory cell $M_t$ is made by the $\tan H$ layer and is represented by:

$M_t=\tan H\left[w^M\left(F_t, Y_{t-1}\right)+c^M\right]$              (4)

In Eq. (4), $\tan H$ permits LSTM to remove or add information in the final input. The information gateway chooses the memory unit, and the forget gate is used to delete or hold data for creating the last memory.

$M_t=F_G * M_{t-1}+I_G * M_t$              (5)

In Eq. (5), $M_t$ signifies memory unit at existing timestamp (t).

$Y_t=O_G * \tan H\left(M_t\right)$              (6)

In Eq. (6), * denotes element‐wise multiplication, $y_t$ indicates the output attained by the softmax resultant layer to obtain the output predictive in the existing blocks. Lastly, the loss function is estimated by selecting the MSE as the error calculation:

$Loss=\sum_{t=1}^N\left(Y_t-T_t\right)^2$              (7)

In Eq. (7), Tt shows desired outcome. N indicates prediction made in instance of n data point. Figure 3 demonstrates infrastructure of LSTM.

3.jpg

Figure 3. LSTM architecture

3.2 Stage II: Hyperparameter tuning process

For hyperparameter tuning of LSTM method, SIOPA used. An OPA is a current metaheuristic optimization model that sketched inspiration from natural hunting strategy of orcas (killer whales) [20]. It stimulates the hunting behavior of orca to resolve the problem of the optimization algorithm. In the OPA, initialize the population of virtual orca, and each orca signifies the best possible solution to the optimization problems.

Driving phase

Current advancement has introduced the incorporation of social interaction, acceleration mechanism, memory, and learning during the driving stage. This increases the capabilities and performance of the model. The acceleration mechanism has been integrated to enhance the exploitation and exploration. Orcas collectively improve their intelligence by exchanging and communicating data about the best possible solution, which facilitates the convergence towards the best solution.

i) Acceleration

The acceleration mechanism has been integrated to improve exploitation and exploration capabilities. This mechanism enables orcas to finetune movement speed dynamically according to solution quality, leading to effective navigation of searching space. Sequentially, this enables better exploration of promising outcomes and exploitation of potential areas. During the chase phase, the velocity update equation can be used as follows:

$Vtchase, 1, i=a *(d * x tbest -F *(b * M t+c * x t i))+w * Vtchase,1, i-1$                 (8)

$Vtchase, 2, i=e *(x t best -x t i)+w * Vtchase , 2, i-1$               (9)

Now, the acceleration or weight feature w controls impact of previous velocity on existing velocity.

ii) Memory and learning

Here, memory and learning mechanism incorporated to allow orcas to exploit previous knowledge. This allows for making appropriate decisions in future iterations and retaining the memory of successful outcomes, gradually increasing performance of model. In chase stage, velocity update equation is adapted to integrate memory and learning as follows:

$Vtchase , 1, i=a *(d * x tbest -F *(b * M t+c * x t i))+w * Vtchase , 1, i-1+m *(Xbest -x t i)$                (10)

In Eq. (10), weight or learning factor m defines impact of prior optimum location (Xbest) on the existing velocity. Orcas could successfully learn from previous experience and accordingly adapt the movement by adjusting these weights. 

iii) Social interaction

During driving process, social interactions amongst orcas established to enable collaboration as well as data sharing. This f to cooperate and exchange valuable information, resulting in better exploitation and exploration of the searching space. During chase phase, velocity updating formula is modified to integrate social interaction:

$Vtchase, 1, i=a *(d * xtbest -F *(b * M t+c * x t i))+w * Vtchase, 1, i-1+s *(Xsocial -x t i)$               (11)

The social or weight factor s defines the impact of social information, Xsocial, on existing velocity.

Encircling phase

A “BubbleNet” mechanism is employed in encircling stage, to enable encircling behaviors of orcas. The orca works together in encircling stage, to encircle target performance by forming a virtual "net" around it.  The orca creates a cooperative force that successfully encircles target outcome by coordinating the action.

i) BubbleNet formation

The orcas exploit a BubbleNet development in the encircling stage, based on the supportive hunting strategy used by orcas. The BubbleNet development assists in concentrating and corraling the target solution in a certain region, which improves efficiency of the collective hunting. The location updating formula for the 3rd chasing method, integrating the BubbleNet creation shown as follows:

$xtchase , 3, i, k=x t d 1, k+u *(x t d 2, k-x t d 3, k)+b * B t$            (12)

$B t=\frac{\sum N n(x t b e s t-x t i)}{N n}$              (13)

After choosing the third chasing technique, xtchase, 3,i,k shows the updated location with BubbleNet development. The BubbleNet creation can be obtained by adding the weighted sum of differences among the past best location (xtbest) and the existing location of orcas (xti). The weight b defines the impact of BubbleNet formation on the movement and allows for coordination of the position for creating the virtual net. The differences between the historical best position and the existing positions for each orca are divided and summed by the overall number of orcas (Nn) to compute the BubbleNet force (Bt).

ii) BubbleNet position modifications

In encircling stage, orca location is modified according to BubbleNet formation, taking fitness function into account. The location updating equation is given below:

$xtchase,i=x tchase, i\,\, i f \,\,f(x t c h a s e, i)<f(x t i)$              (14)

After integrating the BubbleNet formation (xtchase, i), if the fitness values of the location are superior to the fitness values of the existing location (xti), the location remains the same. This ensures that the orca maintains the position if BubbleNet formation does not result in an enhancement in fitness value.

$xtchase,i=x ti \,\,if \,\,f(x t c h a s e, i) \geq f(x t i)$             (15)

After integrating BubbleNet development, if the fitness value of the location is equal or not superior to the fitness values of the existing location, then the location can be upgraded to that existing location. This prevents the orca from moving towards a lesser optimum location.

iii) Adaptive attack speed

During attacking stage, adaptive attack speed is established to change movement rapidity dynamically based on vicinity of the prey and the existing iteration. These facilities improve the probability of catching the prey and improve their attack strategy. S(t), denotes the adaptive attack speed function and evaluates the correct attack speed according to factors including convergence criteria, distance to the prey, and prey movement. During the attacking phase, the velocity update for the orca is expressed as follows:

$Vtattack, 1, i=\frac{(x t 1+x t 2+x t 3+x t 4)}{4}-x tchase,i * S(t)$                (16)

$Vtattack,2, i=\frac{{ (xtchase,d1+xtchase,d2+xtchase,d3) }}{3}-x t i * S(t)$                (17)

Now, velocity upgrade for orca in 1^st and 2^nd attacking strategies are, Vtattack, 1, i and Vtattack, 2, i. According to the orca position, the updates are calculated and their chase target, considering the adaptive attack speed as (t). During the attacking phase, the updated location of the orca is defined as follows:

$xtattack, i= xtchase, i+g 1 * Vtattack, 1, i+g 2 * Vtattack, 2, i$                  (18)

In Eq. (18), the upgrade place of the orcas is xtattack, i considering the chase location, the velocity update, and the weighted g1 and g2. Both weights controlling the impact of the velocity update on the movement, enabling to finetune the attack strategy.

The fitness optimum is an important feature in SIOPA algorithm. An encoded performance has been organized to assess superior efficiency of candidate results. Currently, accuracy value is a main form used to project an FF.