A Novel Federated Incremental Deep Learning Framework for Zero-Day Cyber Attack Detection

Abstract

Zero-day cyber-attacks pose a substantial hazard to modern networked systems due to their unknown signatures and ever-changing attack patterns. The centralized and static learning models that are the mainstay of traditional intrusion detection systems restrict their flexibility, scalability, and adherence to data privacy regulations. This research suggests a novel federated incremental deep learning framework for zero-day cyber-attack detection to overcome these difficulties. Without sharing raw data, the suggested architecture allows remote clients to cooperatively improve a global detection model using federated learning while performing incremental deep learning on locally generated network traffic. The framework successfully manages idea drift and instantly adjusts to new assault patterns by constantly changing model parameters. In comparison to centralized and static federated learning approaches, the suggested framework achieves higher detection accuracy, lower false positive rates, and improved adaptability while guaranteeing data privacy and scalability in distributed environments, according to extensive experiments carried out on benchmark intrusion detection datasets.