Federated TSRN-Enabled GANs for Effective Cyber Attack Detection in Edge Computing

Edge Computing (EC) processes data locally at edge servers, reducing latency, conserving bandwidth, enhancing privacy, and managing large data volumes independently of central servers, making it critical for applications like autonomous vehicles and healthcare. However, EC is vulnerable to cyber attacks, as edge nodes can compromise edge servers and exploit communication between them. Traditional Machine Learning (ML) methods, which require centralizing communication (activity) data, are ineffective for privacy-sensitive EC scenarios. Federated Learning (FL) offers a distributed ML approach for Federated Cyber-Attack Detection (FedCAD), preserving data privacy by training a shared model locally at each edge server, with a central server aggregating updates to form a global model. However, FedCAD faces challenges with unlabeled and non-Independent and non-Identically Distributed (non-IID) activity data. Existing Generative Adversarial Networks (GANs) methods for synthetic data generation do not fully address these issues. To overcome these challenges, we propose FedTSRGNet, which integrates novel Temporal Sequential Recurrent Network (TSRN)-enabled GANs with FL to generate realistic sequential data by capturing complex temporal patterns in local datasets. Additionally, we introduce a reformulated loss function to optimize TSRN-enabled GAN training within the FL paradigm, improving convergence and ensuring the synthetic data closely mimics real data. Theoretical analysis and extensive experiments on two benchmark datasets demonstrate that FedTSRGNet outperforms four state-of-the-art methods by achieving up to 89% accuracy for unlabeled non-IID and 95% accuracy in labeled IID distributions, representing a significant improvement over comparative methods.