Privacy-Preserving Federated Learning Framework Using Dynamic Privacy for Network Intrusion Detection in Cloud Computing

Abstract

Cloud computing has expanded rapidly with the adoption of distributed applications and network-based services, generating large-scale, heterogeneous traffic that requires efficient, privacy-preserving intrusion detection. Conventional intrusion detection systems suffer from limitations such as centralised data dependency, privacy leakage, restricted scalability, and reduced adaptability in distributed environments. To address these challenges, this work proposes Privacy-Preserving Federated Learning with Dynamic Client Selection Roulette Wheel (PPFL-DCSRW). The CSE-CIC-IDS2018 dataset is preprocessed using Robust Maximum Correntropy Kalman Filter (RMCKF) to suppress noise and reduce data bias. Feature extraction is performed using the Fractional Adaptive Superlet Transform (FASLT), capturing temporal spread, adaptive frequency patterns, and localised spectral characteristics of network traffic. The extracted features are trained using Fed-DCSRW, in which clients are grouped by similarity in local learning behaviour while preventing raw data exchange, thereby enabling privacy-preserving collaborative intrusion detection across distributed cloud environments. The framework is implemented using Python and evaluated in a distributed cloud environment containing heterogeneous clients under client dropout, latency, and selective denial-of-service conditions, achieving an average detection rate of 98.13% and a false positive rate of 1.46%, demonstrating improved communication efficiency, faster convergence, and reduced recovery time compared with baseline federated methods.