Privacy-Preserving Intrusion Detection Using Federated Learning: Performance Evaluation and Cyber Law Implications for Data Protection and AI Accountability

This study investigates the use of federated learning for privacy-preserving intrusion detection systems and examines its implications from a cyber law perspective, particularly in relation to data protection and AI accountability. A comparative analysis was conducted between centralized machine learning models, including Random Forest and Extra Trees, and a federated Multilayer Perceptron (MLP) enhanced with the FedProx algorithm under both IID and non-IID data distributions. The results show that the centralized Random Forest model achieved the highest performance, while the federated model under IID conditions demonstrated comparable results, indicating that federated learning can maintain high detection capability without requiring centralized data collection. However, performance degradation was observed under non-IID settings, highlighting the impact of data heterogeneity on federated systems. From a legal perspective, federated learning supports key data protection principles such as data minimization and privacy by design by keeping sensitive data on local devices. Nevertheless, challenges related to model security, limited explainability, and potential information leakage remain, raising concerns regarding AI accountability and regulatory compliance. This study concludes that federated learning offers a promising approach for secure and privacy-aware intrusion detection, but requires additional safeguards to ensure both technical robustness and alignment with evolving cyber law frameworks.