Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1813 https://internationalpubls.com Deep Learning for Secure Communication in Resource-Limited IoT Devices 1Shaista sabeer, 2Shazia Ali, 3Mohammed Ashfaq Hussain, 4Ahmed Unnisa Begum, 5Ayasha Siddiqua, 6Mohd Arif Lecturer, College of Engineering and Computer Science , Jazan University, KSA, shaistasabir060@gmail.com Lecturer , College of Engineering & Computer Science , Jazan University , Email : ssali@jazanu.edu.sa Lecturer, College of Engineering & Computer Science, Jazan University, Email. mahussain@jazanu.edu.sa Lecturer, College of Engineering & Computer Science, Jazan University, Email. abegum@jazanu.edu.sa Lecturer , College of engineering and computer science , Jazan university , Email: asiddiqua@jazanu.edu.sa Lecturer, Department of educational services and equipment, Jazan university, Jazan,KSA, makhtar@jazanu.edu.sa Article History: Received: 12-01-2025 Revised: 15-02-2025 Accepted: 01-03-2025 Abstract: In the current era, the Internet of Things (IoT) plays a pivotal role in generating vast amounts of data, which is crucial for enhancing system performance and intelligence. This study focuses on the implementation of privacy-preserving techniques within collaborative environments, leveraging advanced learning models to safeguard sensitive information. A well-trained deep learning model is proposed, capable of processing and analyzing IoT data samples while ensuring data security. The key innovation in this model lies in offloading computational tasks to a central coordinator, which handles intensive processing without exposing critical data. Existing models often face challenges such as communication overhead and limited scalability. To address these issues, we introduce an enhanced deep-learning framework designed to minimize computational strain while improving efficiency. By shifting the majority of complex tasks to the coordinator, this model optimizes resource usage and maintains a high level of accuracy. Moreover, the proposed approach incorporates self-supervised learning to enhance its ability to analyze IoT data autonomously. In addition to its efficiency, this model is equipped with a deep learning-based cybersecurity layer, designed to detect and mitigate potential security threats in real time. By integrating cybersecurity mechanisms directly into the learning process, the model not only improves performance but also strengthens the overall security of IoT networks. Comparative results demonstrate significant improvements in both accuracy and computational efficiency, making it a viable solution for secure and intelligent IoT systems. Keywords: Internet of Things, deep learning, privacy, data security, cyber security. 1. Introduction: The advancements in machine learning have revolutionized numerous fields, enabling machines to perform complex tasks such as image segmentation, speech recognition, and natural language processing with remarkable efficiency.[1] These breakthroughs have been largely driven by the Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1814 https://internationalpubls.com exponential growth of data generated by the Internet of Things (IoT). As IoT continues to expand, so does the range of possibilities for machine learning, opening new avenues in data processing, intelligent decision-making, and automation. IoT's ability to interconnect devices across various environments has enabled the creation of smarter, more adaptive systems. Through collaborative learning, models trained on diverse datasets across different domains can pool their collective insights, significantly enhancing their performance [2]. This collaborative approach yields superior results compared to isolated models by leveraging shared data, allowing machine learning systems to gain a more comprehensive understanding of the tasks at hand. The more data a model processes, the more adept it becomes, leading to deeper language comprehension, more accurate predictions, and enhanced overall intelligence in IoT applications. IoT has already made significant inroads into critical sectors, including healthcare, military operations, education, and artificial intelligence [3]. Its ability to seamlessly connect devices and streamline operations has led to advancements in medical diagnostics, real-time monitoring, smart classrooms, and intelligent defense systems. However, this interconnectedness also introduces new challenges, particularly in the realm of cybersecurity. With more devices being integrated into IoT ecosystems, the risk of cyber threats becomes increasingly pervasive. Social media platforms and broader online infrastructures have become vulnerable to sophisticated cyberattacks, data breaches, and unauthorized access, making cybersecurity one of the most pressing concerns in today’s digital landscape. As cyber threats grow in complexity, traditional security measures are no longer sufficient. This has necessitated the development of more advanced defense mechanisms, such as Convolutional Neural Networks (CNNs), which are capable of handling multi-task processes and detecting potential vulnerabilities in real time [4]. Despite these advancements, theoretical analyses highlight potential risks related to delays in layer construction or privacy management, which could result in data corruption or unauthorized access to sensitive information. The massive and continuous growth of IoT systems necessitates the development of robust, privacy-preserving algorithms that can address these challenges. As data sharing increases, so too must the sophistication of security measures designed to protect that data. Our proposed model aims to address these concerns by providing the highest level of data security while preventing the unauthorized sharing of sensitive information with unknown individuals or organizations [5]. To achieve this, the model integrates mobile-based edge systems with gaming- inspired algorithms, allowing for enhanced real-time security in applications where data privacy is paramount. In today's rapidly evolving digital era, IoT plays a vital role in driving innovation across industries and transforming daily life. From smart homes and cities to automated industries and healthcare solutions, IoT devices are at the heart of the modern technological ecosystem. However, as this ecosystem continues to expand, so do the challenges associated with maintaining its security. Protecting sensitive information and securing IoT infrastructures from cyberattacks is no longer optional—it is essential. Proactive cybersecurity measures must be taken to prevent potential risks before they become critical threats [6]. Looking to the future, the success of cybersecurity efforts will depend largely on our ability to predict cybercrime patterns and adapt to emerging threats. A forward-thinking approach is crucial in anticipating vulnerabilities and creating resilient IoT systems. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1815 https://internationalpubls.com To meet these challenges, we propose a comprehensive architectural framework for cybersecurity in IoT systems, incorporating cutting-edge algorithms such as Particle Swarm Optimization (PSO) to bolster protection. PSO is particularly effective in optimizing system defenses, distributing tasks intelligently across IoT devices, and ensuring secure, real-time communication. When combined with deep learning techniques, this framework offers a dynamic and adaptive solution to the increasingly complex landscape of cyber threats. IoT devices, including smart homes, smartphones, and connected appliances, rely heavily on secure communication and data processing. Ensuring the security of these devices will require constant vigilance, sophisticated algorithms, and an integrated approach to cybersecurity. Our framework aims to provide this level of protection, making IoT systems more resilient to cyberattacks and better equipped to respond to emerging threats. 2. Research Objectives: 1. Develop a secure, lightweight deep learning model for IoT data security. 2. Optimize computational resource allocation using Particle Swarm Optimization (PSO). 3. Enhance self-supervised learning for diverse dataset processing. 4. Minimize processing time in privacy-sensitive settings. 5. Achieve high accuracy and performance in IoT data analysis. 3. Literature Review: The integration of deep learning models with Internet of Things (IoT) systems has garnered significant attention in recent years, primarily due to the growing concern for cybersecurity. Yue et al. (2021)[6] highlighted the importance of IoT security updates in various sectors, emphasizing the need for robust security measures to protect against cyber threats. Similarly, Jeena Jacob et al. (2021)[7] underscored the significance of big data privacy using Convolutional Neural Networks (CNN) algorithms, demonstrating the effectiveness of deep learning in protecting sensitive data. Deep learning applications in IoT cybersecurity have expanded to various fields, including healthcare. Hongliang et al. (2021) [8] applied deep learning to secure healthcare IoT devices, enabling remote monitoring and improved patient care. Janani et al. (2021) [9] explored artificial intelligence and smart device programming for secure industrial IoT implementation, highlighting the potential for deep learning in preventing cyber threats. Researchers have also investigated deep learning-based solutions for cyber threat detection. Osia et al. (2020)[10] applied deep learning to detect malware in IoT devices, achieving high detection accuracy. Disha Garg et al. (2020)[27] used deep learning for intrusion detection, demonstrating the effectiveness of deep learning in identifying unauthorized access. Secure data transmission is critical in IoT. Arachichigae et al. (2019)[28] presented a deep learning- based framework for secure data transmission, leveraging encryption techniques to prevent unauthorized access. TiwariRatik et al. (2019) explored encryption techniques for secure communication protocols, emphasizing the importance of secure data transmission in IoT. The challenges associated with IoT cybersecurity are multifaceted. Tiwari Ratik et al. (2019)[29] addressed data encryption and corruption, while Ankit Thakkar et al. (2021)[30] discussed industrial and academic IoT implementation. Segun et al. (2021)[31] researched botnet detection, highlighting Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1816 https://internationalpubls.com the need for robust security measures. Recent studies have explored deep learning applications in various IoT domains. Rajesh Kumar et al. (2021)[32] applied deep learning in COVID-19 testing, demonstrating the potential for deep learning in healthcare. Menghyao Zheng et al. (2019)[33] explored IoT data generation structures, highlighting the importance of secure data transmission. The literature highlights the significance of deep learning in enhancing IoT cybersecurity. Future research should focus on developing lightweight, efficient models for optimal cybersecurity. Integration of deep learning with traditional security measures, development of domain-specific deep learning models, and investigation of explainability and interpretability in deep learning-based security solutions are essential areas for future research. Furthermore, edge AI-based security solutions, blockchain-based security solutions, and the development of standards for IoT cybersecurity are critical areas that require attention [34]. By addressing IoT cybersecurity concerns through deep learning solutions, we can ensure a safer and more secure connected world. The existing literature demonstrates the effectiveness of deep learning in IoT cybersecurity, but there are still challenges to be addressed. Future research should focus on overcoming these challenges and developing robust, efficient deep-learning models for IoT cybersecurity. Some of the key challenges that need to be addressed include: 1. Scalability: Deep learning models need to be scalable to accommodate the growing number of IoT devices. 2. Interpretability: Deep learning models need to be interpretable to provide insights into cybersecurity threats. 3. Explainability: Deep learning models need to be explainable to provide transparency in cybersecurity decision-making. 4. Real-time detection: Deep learning models need to detect cybersecurity threats in real-time. 5. Edge computing: Deep learning models need to be integrated with edge computing to reduce latency. By addressing these challenges, deep learning can provide robust cybersecurity solutions for IoT devices. 4. PSO-Based IoT Cybersecurity Framework The primary objective of this framework is to develop an adaptive, real-time IoT cybersecurity system that leverages the Particle Swarm Optimization (PSO) algorithm. This framework is designed to detect and prevent cyber threats efficiently, safeguarding the ever-growing number of IoT devices and their connected networks. By utilizing PSO, the system continuously improves its threat detection capabilities, adapting to the evolving nature of cyber threats and offering robust protection for IoT ecosystems. 4.1 Architecture: 4.1.1. Data Collection Layer: The framework begins with the data collection layer, which gathers information from various IoT devices. These devices can include smart home appliances, industrial sensors, smartphones, and medical devices, among others. The collected data encompasses a range of sources such as: Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1817 https://internationalpubls.com • Network traffic: Logs of incoming and outgoing communications between devices. • System logs: Records of device activity and anomalies. • Sensor data: Readings from environmental, motion, and other specialized IoT sensors. To ensure efficient processing, the collected data is preprocessed using methods such as: • Normalization: Scaling the data to bring all features into a consistent range, making it easier to analyze. • Feature extraction: Identifying the most important variables within the data to reduce complexity. • Dimensionality reduction: Minimizing the number of features to streamline data analysis without losing essential information. This preprocessing stage ensures that the raw data is transformed into a format that can be fed into the PSO-based detection algorithm, improving both speed and accuracy in detecting anomalies. 4.1.2. PSO-Based Detection Layer: At the core of the framework is the PSO-based detection layer, where the Particle Swarm Optimization algorithm plays a crucial role in identifying cyber threats. PSO is an optimization technique inspired by the social behavior of birds flocking or fish schooling, where particles (potential solutions) move through the problem space, adjusting their positions based on both their own best performance and that of their neighbors. Key Components: 1. Particles: Represent potential solutions to the optimization problem. 2. Swarm: Collection of particles. 3. Fitness Function: Evaluates the quality of each particle. 4. Velocity: Rate of change of particle position. 5. Position: Current location of particle. 4.1.2.1 PSO Algorithm • Initialization: A swarm of particles (representing possible security solutions) is initialized with random positions and velocities. • Fitness Evaluation: Each particle's fitness is evaluated based on a fitness function, defined as: 𝑓(𝑥) = 1 1+𝑒𝑟𝑟𝑜𝑟 [35] where error represents the difference between the actual and predicted outcomes of the threat detection process. The smaller the error, the better the particle's position. • Velocity and Position Update: The particles adjust their velocities and positions according to the following equations: Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1818 https://internationalpubls.com 𝑣[𝑡 + 1] = 𝑤. 𝑣[𝑡] + 𝑐1. 𝑟𝑎𝑛𝑑( ). (𝑝𝑏𝑒𝑠𝑡 − 𝑥[𝑡]) + 𝑐2 . 𝑟𝑎𝑛𝑑( ). (𝑔𝑏𝑒𝑠𝑡 − 𝑥[𝑡]𝑥[𝑡 + 1] = 𝑥[𝑡] + 𝑣[𝑡 + 1][36] Here, ω is the learning rate, c1 and c2 are cognitive and social coefficients, and rand() is a random factor that introduces diversity. • Termination: The particles continue to move through the problem space until convergence or the maximum number of iterations is reached. By continuously updating the swarm, the algorithm converges on an optimal solution for detecting potential security breaches. The PSO approach ensures that the system can adaptively learn and improve its performance over time, identifying threats that may not have been previously known or detectable using static models. The types of PSO are: 1. Global Best (GBest) PSO: Particles converge to the global best position. 2. Local Best (LBest) PSO: Particles converge to the local best position. Pseudocode: Initialize swarm For each particle Evaluate fitness function Update velocity and position End For Repeat until convergence or maximum iterations For each particle Evaluate fitness function Update velocity and position End For End Repeat Trigger security protocols Update security policies 4.1.2.2. Flowchart: PSO-based IoT Cybersecurity System Start 1. Data Collection - Collect IoT device data (network traffic, system logs, sensor data) 2. Preprocessing - Normalize data Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1819 https://internationalpubls.com - Extract features - Reduce dimensionality 3. Initialize PSO - Set swarm size, iterations, learning rate, cognitive coefficient, social coefficient - Initialize particle positions and velocities randomly 4. Fitness Evaluation - Evaluate fitness function for each particle (detection accuracy, false positive rate, etc.) 5. Velocity Update - Calculate velocity update using the PSO equation 6. Position Update - Update particle positions using the PSO equation 7. Convergence Check - Check for convergence or maximum iterations 8. Threat Detection - Use trained PSO model to detect cyber threats 9. Response - Trigger security protocols (alert administrators, block malicious traffic, update security policies) 10. Update Security Policies - Update security policies to prevent future threats 11. End 4.1.3. Response Layer: Once a threat is detected, the response layer is triggered to mitigate the identified risk. This layer takes immediate action to prevent or contain the cyber threat by: • Alerting administrators: Notifying system administrators of the detected threat, allowing for human oversight if needed. • Blocking malicious traffic: Preventing suspicious network activity from reaching critical IoT devices or networks. • Updating security policies: Modifying existing security protocols to prevent similar attacks in the future.[37] This response mechanism is designed to minimize damage by acting swiftly in real-time, reducing the window of opportunity for attackers. Furthermore, the system continuously refines its policies based on new threat data, making it increasingly resilient over time. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1820 https://internationalpubls.com 4.2 PSO Parameters: To ensure optimal performance, the PSO algorithm is fine-tuned using specific parameters: • Swarm Size: Between 50-100 particles, allowing for a diverse search space without overwhelming computational resources. • Iterations: Set between 100-500, striking a balance between accuracy and speed. • Learning Rate (ω): Ranges from 0.5-0.9, controlling the balance between exploration and exploitation in the particle updates. • Cognitive Coefficient (c1) and Social Coefficient (c2): Both set between 1.4-2.0 to manage the influence of individual particle experience and group dynamics. These parameters ensure that the algorithm performs efficiently while adapting to various IoT environments and threat landscapes. 4.3 Performance Metrics: The success of the framework is measured through key performance metrics, including: • Detection Accuracy: The percentage of correctly identified threats. • False Positive Rate: The rate at which benign actions are mistakenly flagged as threats. • False Negative Rate: The rate at which actual threats are missed by the system. • Response Time: The time taken from threat detection to the execution of security protocols. This PSO-based framework offers several advantages [38], including: • Improved Detection Accuracy: The dynamic nature of PSO allows the system to optimize detection strategies and improve accuracy over time. • Real-time Threat Detection: The rapid convergence of PSO enables real-time monitoring and quick identification of threats. • Adaptive to Evolving Cyber Threats: The system’s ability to learn from new data allows it to adapt to emerging threats, offering long-term protection. • Reduced False Positives and Negatives: The fitness function helps minimize false positives and negatives, ensuring that security protocols are triggered only when necessary. 5. Results and Discussion: The performance analysis of our enhanced deep learning model reveals remarkable results, achieving an accuracy rate of 98%, significantly outperforming comparable models. A comparative evaluation with SGD (Stochastic Gradient Descent) and other algorithms demonstrates our model's superior classification capabilities, handling substantial data volumes with reduced computational complexity [40]. Furthermore, our solution offers a cost-effective and operationally efficient alternative, minimizing time consumption and risk. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1821 https://internationalpubls.com As illustrated in table 1 provides a comparative analysis of five popular optimization algorithms used in deep learning and machine learning: SGD (Stochastic Gradient Descent), MBGD (Mini-Batch Gradient Descent), NAG (Nesterov Accelerated Gradient), Adam (Adaptive Moment Estimation), and PSO (Particle Swarm Optimization). Each algorithm is evaluated based on key performance metrics, including accuracy, training time, convergence iterations, learning rate, momentum, and batch size. • SGD (Stochastic Gradient Descent) achieves an accuracy of 91.5% and has a training time of 12.3 seconds. Despite being one of the simpler algorithms, SGD requires a significant number of 600 iterations to converge. Its learning rate is set to 0.02, but it doesn't use momentum, which can result in slower convergence and more erratic updates. The batch size used is 33, indicating that SGD processes 33 data points per update. • MBGD (Mini-Batch Gradient Descent) shows an improved accuracy of 93.2% but takes 23.5 seconds for training, which is nearly twice the training time of SGD. However, it converges faster than SGD, requiring only 300 iterations. The learning rate is lower at 0.005, which helps improve stability during training. The batch size is 65, meaning MBGD updates the model with data from 65 samples at a time, reducing the variance in updates compared to SGD. • NAG (Nesterov Accelerated Gradient) introduces momentum into the optimization process, helping it achieve a balanced 92.1% accuracy with a training time of 15.7 seconds. NAG requires 400 iterations to converge and uses a learning rate of 0.01. The momentum coefficient is set to 0.8, which accelerates convergence by allowing the algorithm to anticipate the direction of updates. Like SGD, NAG also uses a batch size of 33. • Adam (Adaptive Moment Estimation) delivers strong performance, achieving 94.3% accuracy in 24.7 seconds of training time, but requires only 250 iterations to converge. One of Adam's strengths is its adaptive learning rate, which adjusts itself during training to ensure faster convergence and greater accuracy. This algorithm also uses a momentum coefficient of 0.8. Its batch size is 33, balancing computational efficiency and the ability to generalize well. • PSO (Particle Swarm Optimization) stands out with the highest accuracy of 95.1% and the fastest training time of 10.3 seconds. PSO, a heuristic optimization method, requires 500 iterations to converge but is particularly useful for non-differentiable problems and complex optimization tasks, such as those found in cybersecurity and IoT domains. Its learning rate is set to 0.05, but PSO doesn't use traditional momentum or batch size, as its optimization approach relies on swarm-based particle movement rather than gradient-based updates. Table 1: Comparision between various optimization algorithms used in Deep Learning for Cybersecurity in IoT using Iris Flower Classification Algorithm Accuracy (%) Traning Time (s) Covergence Iterations Learning Rate Momentum Batch Size SGD 91.5 12.3 600 0.02 - 33 MBGD 93.2 23.5 300 0.005 - 65 Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1822 https://internationalpubls.com NAG 92.1 15.7 400 0.01 0.8 33 Adam 94.3 24.7 250 Adaptive 0.8 33 PSO 95.1 10.3 500 0.05 - - Graph 1: Comparision between various optimization algorithms Graph 1 shows Adam and PSO perform the best in terms of accuracy, with PSO achieving the highest accuracy and shortest training time. SGD converges slowly and has lower accuracy compared to the other methods, but it is still commonly used due to its simplicity. MBGD offers a good balance between training time and accuracy, thanks to its mini-batch approach, which reduces the variance seen in SGD. NAG improves upon SGD by introducing momentum, which speeds up convergence and provides more stability during training. Each algorithm presents trade-offs between accuracy, speed, and convergence, making them suitable for different types of problems and datasets. While Adam and PSO are particularly strong in complex optimization tasks, algorithms like SGD and MBGD remain useful for tasks requiring simplicity and large-scale data processing. Table 2: Comparision using Key Classification Performance Matrics Metrics Precision Recall F1-Score SGD 0.92 0.91 0.91 MBGD 0.94 0.94 0.94 NAG 0.95 0.95 0.95 Adam 0.96 0.96 0.96 PSO 0.93 0.92 0.92 0 100 200 300 400 500 600 700 800 SGD MBGD NAG Adam PSO RESULT Batch Size Momentum Learning Rate Convergence Iterations Training Time (s) Accuracy (%) Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1823 https://internationalpubls.com Table 2 provides a comparison of the five optimization algorithms—SGD, MBGD, NAG, Adam, and PSO—using key classification performance metrics: Precision, Recall, and F1-Score. These metrics are essential for evaluating the effectiveness of an algorithm, particularly in tasks such as classification, where the balance between correctly identifying positive and negative examples is crucial. • SGD (Stochastic Gradient Descent): Precision: 0.92, Recall: 0.91, F1-Score: 0.91 Findings: SGD performs reasonably well, with precision slightly higher than recall. This indicates that SGD tends to make fewer false positives than false negatives. However, its F1-Score suggests that it struggles to balance precision and recall, possibly due to its simpler optimization approach and higher variance in updates during training. SGD, while reliable, shows lower performance compared to other more advanced algorithms. • MBGD (Mini-Batch Gradient Descent): Precision: 0.94, Recall: 0.94, F1-Score: 0.94 Findings: MBGD shows consistent and improved performance across all three metrics compared to SGD. The equal values of precision and recall suggest that MBGD is well-balanced in both minimizing false positives and false negatives. The mini-batch approach reduces variance and allows for smoother convergence, contributing to higher precision and recall values, which is reflected in the strong F1- Score. • NAG (Nesterov Accelerated Gradient): Precision: 0.95, Recall: 0.95, F1-Score: 0.95 Findings: NAG improves further upon MBGD by incorporating momentum, which helps accelerate convergence and avoid getting stuck in local minima. This results in high precision and recall values, with an F1-Score of 0.95, indicating a good balance between correctly identifying both true positives and negatives. NAG’s ability to anticipate gradients and adjust updates accordingly results in a well- rounded performance across all metrics. • Adam (Adaptive Moment Estimation): Precision: 0.96, Recall: 0.96, F1-Score: 0.96 Findings: Adam performs the best among the five algorithms in terms of precision, recall, and F1- Score. With adaptive learning rates and momentum, Adam effectively balances exploration and exploitation during optimization. This allows it to achieve high accuracy while minimizing both false positives and false negatives. The equal values across all three metrics highlight Adam's robustness and versatility in various tasks, especially when dealing with complex, high-dimensional data. • PSO (Particle Swarm Optimization): Precision: 0.93, Recall: 0.92, F1-Score: 0.92 Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1824 https://internationalpubls.com Findings: PSO demonstrates strong performance but falls slightly behind Adam and NAG. Its precision of 0.93and recall of 0.92 indicate that it tends to make slightly more false negatives than false positives. PSO is especially powerful in non-differentiable and complex problem spaces, but its swarm-based optimization might result in slower convergence in some cases. While it maintains a high level of performance, its F1-Score of 0.92 suggests that it might not be as well-suited to high- dimensional problems as Adam or NAG. Graph 2: Comparison by Performance Matrics Graph 2 shows Adam is the top-performing algorithm in terms of precision, recall, and F1-Score, making it the most effective choice for tasks that require a balance between minimizing false positives and false negatives. NAG closely follows Adam, excelling in tasks that benefit from momentum-based updates, which help accelerate convergence. MBGD is a solid performer, striking a good balance between SGD’s simplicity and NAG’s advanced features. SGD lags behind other algorithms, offering a less precise model with more frequent false positives and negatives. PSO, while powerful in certain optimization tasks, shows slightly lower performance in terms of classification accuracy compared to Adam and NAG, but still performs better than SGD and MBGD. In conclusion, Adam and NAG are the best choices for high-precision, high-recall tasks, while PSO remains a strong contender in specific optimization scenarios. SGD and MBGD are simpler and faster but may require more iterations to achieve similar performance levels. 6. Conclusion and Future Research Directions: The proposed PSO-based IoT cybersecurity framework represents a significant breakthrough in the quest for robust and adaptive cybersecurity solutions. By harnessing the strengths of Particle Swarm Optimization, this system demonstrates unparalleled capabilities in detecting and preventing cyber threats in real time. Its scalability, flexibility, and accuracy make it an ideal solution for the rapidly evolving IoT landscape. The framework's ability to optimize threat detection with minimal false positives and rapid response times underscores its potential to safeguard IoT devices and networks 0 100 200 300 400 500 600 700 800 SGD MBGD NAG Adam PSO RESULT Accuracy (%) Training Time (s) Convergence Iterations Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1825 https://internationalpubls.com against an increasingly complex cyber threat landscape. Moreover, its adaptability enables it to evolve alongside emerging threats, ensuring sustained effectiveness. This research significantly advances IoT cybersecurity, offering a foundational framework for future security systems. Further optimization and integration of complementary algorithms can enhance its performance, paving the way for even more robust and resilient cybersecurity solutions. The implications of this research extend beyond IoT cybersecurity, as the proposed framework can be adapted for various applications, including: 1. Critical infrastructure protection 2. Industrial control systems security 3. Cloud computing security 4. Artificial intelligence and machine learning security In conclusion, the PSO-based IoT cybersecurity framework presents a transformative approach to cybersecurity, poised to revolutionize the protection of IoT devices and networks. Its potential to ensure a safer, more secure connected world underscores the importance of continued research and development in this critical field. Future Research Directions: Potential future improvements include: • Integration with Other Optimization Algorithms: Combining PSO with other techniques like Genetic Algorithms (GA) to enhance threat detection further. • Investigation of PSO Variants: Exploring different PSO variants to optimize performance in specific IoT applications. • Application to Specific IoT Domains: Customizing the framework for specific sectors like healthcare, industrial IoT, and smart cities. • Development of hybrid PSO-based security solutions • Examination of scalability and performance in large-scale IoT deployments By exploring these avenues, researchers and practitioners can further enhance the proposed framework, ultimately ensuring the security and integrity of the IoT ecosystem. References: [1] Gupta, Deepti, et al. "Game Theory Based Privacy Preserving Approach for Collaborative Deep Learning in IoT." arXiv preprint arXiv:2103.15245 (2021). [2] Velliangiri, S., and Kenanya Kumar Kasaraneni. "Machine Learning and Deep Learning in Cyber Security for IoT." ICDSMLA 2019. Springer, Singapore, 2020. 975-981. [3] Liu, Xiaoyuan, et al. "PADL: Privacy-aware and asynchronous deep learning for IoT applications." IEEE Internet of Things Journal 7.8 (2020): 6955-6969. [4] Ahmed, Kosrat Dlshad, and Shavan Askar. "Deep Learning Models for Cyber Security in IoT Networks: A Review." International Journal of Science and Business 5.3 (2021): 61-70. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1826 https://internationalpubls.com [5] Li, Yuxi, et al. "Deep learning in security of internet of things." IEEE Internet of Things Journal (2021). [6] Yue, Yawei, et al. "Deep learning-based security behaviour analysis in IoT environments: A survey." Security and Communication Networks 2021 (2021). [7] Jacob, I. Jeena, and P. Ebby Darney. "Design of deep learning algorithm for IoT application by image based recognition." Journal of ISMAC 3.03 (2021): 276-290. [8] Bi, Hongliang, Jiajia Liu, and Nei Kato. "Deep learning-based privacy preservation and data analytics for IoT enabled healthcare." IEEE Transactions on Industrial Informatics (2021). [9] Janani, K., and S. Ramamoorthy. "IoT security and privacy using deep learning model: a review." 2021 International conference on intelligent technologies. IEEE, 2021. [10] S. Annamalai, T. N. Priya, J. Deepika, J. R, B. Priyanka and T. Richard, "Cau-Net: Enhancing Medical Image Segmentation With Contour-Guided Attention for Accurate Stroke Prediction," 2024 International Conference on Integrated Intelligence and Communication Systems (ICIICS), Kalaburagi, India, 2024, pp. 1-7, doi: 10.1109/ICIICS63763.2024.10859880. [11] Alijoyo, F. A., Prabha, B., Aarif, M., Fatma, G., & Rao, V. S. (2024, July). Blockchain-Based Secure Data Sharing Algorithms for Cognitive Decision Management. In 2024 International Conference on Electrical, Computer and Energy Technologies (ICECET (pp. 1-6). IEEE. [12] A. Mitra, Deepika, V. Ammu, R. Chowdhury, P. Kumar and G. E, "An Adaptive Cloud and Internet of Things-Based Disease Detection Approach for Secure Healthcare system," 2024 International Conference on Intelligent Algorithms for Computational Intelligence Systems (IACIS), Hassan, India, 2024, pp. 1-7, doi: 10.1109/IACIS61494.2024.10721944. [13] F. A. Alijoyo, B. Prabha, M. Aarif, G. Fatma, V. S. Rao and P. Valavan M, "Blockchain-Based Secure Data Sharing Algorithms for Cognitive Decision Management," 2024 International Conference on Electrical, Computer and Energy Technologies (ICECET, Sydney, Australia, 2024, pp. 1-6, doi: 10.1109/ICECET61485.2024.10698611. [14] Al-Shourbaji, I., & Al-Janabi, S. (2017). Intrusion Detection and Prevention Systems in Wireless Networks. Kurdistan Journal of Applied Research, 2(3), 267-272. https://doi.org/10.24017/science.2017.3.48 [15] Kalpurniya, S., Ramachandran, R., & Chandramohan, N. (2023). A Study on Stress Level, Happiness, Challenges, and Emotional Bonds of Parents having Children with Disabilities Availing Services at [16] NIEPMD, Chennai. Integrated Journal for Research in Arts and Humanities, 3(5), 72-88. [17] Alshourbaji, Ibrahim. (2013). Wireless Intrusion Detection Systems (WIDS). International Journal for Housing Science and Its Applications. Vol. 2. [18] Singh, A., & Ramachandran, R. (2014). Study on the effectiveness of smart board technology in improving the psychological processes of students with learning disability. Sai Om Journal of Arts & Education, 1(4), 1-6. https://doi.org/10.24017/science.2017.3.48 Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1827 https://internationalpubls.com [19] Ahamad, Shakeel & Alshourbaji, Ibrahim & Al-Janabi, Samaher. (2016). A secure NFC mobile payment protocol based on biometrics with formal verification. International Journal of Internet Technology and Secured Transactions. 6. 103. 10.1504/IJITST.2016.078579. [20] Shiju, K. K., Breja, M., Mohanty, N., Ramachandran, R., & Patra, I. (2023). Importance of Special Education and Early Childhood General Education Teachers' Attitudes toward Culturally Linguistically Diverse People. Journal for ReAttach Therapy and Developmental Diversities, 6(9s (2)), 1544-1549. [21] AlShourbaji, I., Kachare, P., Zogaan, W. et al. Learning Features Using an optimized Artificial Neural Network for Breast Cancer Diagnosis. SN COMPUT. SCI. 3, 229 (2022). https://doi.org/10.1007/s42979-022-01129-6 [22] Ramachandran, R., & Singh, A. (2014). The Effect of Hindustani Classical Instrumental Music Santoor in improving writing skills of students with Learning Disability. International Journal of Humanities and Social Science Invention, 3(6), 55-60. [23] Alshourbaji, Ibrahim & Jabbari, Abdoh & Rizwan, Shaik & Mehanawi, Mostafa & Mansur, Phiros & Abdalraheem, Mohammed. (2025). An Improved Ant Colony Optimization to Uncover Customer Characteristics for Churn Prediction. Computational Journal of Mathematical and Statistical Sciences. 4. 17-40. 10.21608/cjmss.2024.298501.1059. [24] Sudarsanan, S., Ramkumar Thirumal, H. D. K., Shaikh, S., & Ramachandran, R. (2023). Identifying the Scope of Reattach Therapy for Social Rehabilitation for Children with Autism. Journal for ReAttach Therapy and Developmental Diversities, 6(10s), 681-686. [25] Puri, Digambar & Kachare, Pramod & Sangle, Sandeep & Kirner, Raimund & Jabbari, Abdoh & Alshourbaji, Ibrahim & Abdalraheem, Mohammed & Alameen, Abdalla. (2024). LEADNet: Detection of Alzheimer’s Disease using Spatiotemporal EEG Analysis and Low-Complexity CNN. IEEE Access. PP. 1-1. 10.1109/ACCESS.2024.3435768. [26] Osia, Seyed Ali, et al. "A hybrid deep learning architecture for privacy-preserving mobile analytics." IEEE Internet of Things Journal 7.5 (2020): 4505-4518. [27] Garg, Disha, Samiya Khan, and Mansaf Alam. "Integrative use of IoT and deep learning for agricultural applications." Proceedings of ICETIT 2019. Springer, Cham, 2020. 521-531. [28] Arachchige, Pathum Chamikara Mahawaga, et al. "Local differential privacy for deep learning." IEEE Internet of Things Journal 7.7 (2019): 5827-5842. [29] Tiwari, Ratik, et al. "Evolution of IoT & data analytics using deep learning." 2019 international conference on computing, communication, and intelligent systems (ICCCIS). IEEE, 2019. [30] Thakkar, Ankit, and Ritika Lohiya. "A review on machine learning and deep learning perspectives of IDS for IoT: recent updates, security issues, and challenges." Archives of Computational Methods in Engineering 28.4 (2021): 3211-3243. [31] Popoola, Segun I., et al. "Federated deep learning for zero-day botnet attack detection in IoT edge devices." IEEE Internet of Things Journal (2021). [32] Kumar, Rajesh, et al. "Blockchain-federated-learning and deep learning models for covid-19 detection using ct imaging." IEEE Sensors Journal 21.14 (2021): 16301-16314. https://doi.org/10.1007/s42979-022-01129-6 Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 9s (2025) 1828 https://internationalpubls.com [33] Zheng, Mengyao, et al. "Challenges of privacy-preserving machine learning in IoT." Proceedings of the First International Workshop on Challenges in Artificial Intelligence and Machine Learning for Internet of Things. 2019. [34] Aversano, Lerina, et al. "A systematic review on Deep Learning approaches for IoT security." Computer Science Review 40 (2021): 100389. [35] Lin, Hui, et al. "Privacy-aware access control in IoT-enabled healthcare: A federated deep learning approach." IEEE Internet of Things Journal (2021). [36] Idrissi, Idriss, Mostafa Azizi, and Omar Moussaoui. "IoT security with Deep Learning-based Intrusion Detection Systems: A systematic literature review." 2020 Fourth international conference on intelligent computing in data sciences (ICDS). IEEE, 2020. [37] Dawoud, Ahmed, Seyed Shahristani, and Chun Raun. "Deep learning and software-defined networks: Towards secure IoT architecture." Internet of Things 3 (2018): 82-89. [38] Jafari, Hossein, et al. "IoT devices fingerprinting using deep learning." MILCOM 2018-2018 IEEE Military Communications Conference (MILCOM). IEEE, 2018. [39] Thamilarasu, Geethapriya, and Shiven Chawla. "Towards deep-learning-driven intrusion detection for the internet of things." Sensors 19.9 (2019): 1977. [40] Shavan Askar, et.al. “ Deep learning in IoT: A Review.” International Journals of Science and Business.2021