Enhancing Cybersecurity Attack Detection: An Asynchronous Federated Learning Approach

We propose an asynchronous federated learning (AFL) framework to address inefficiencies in traditional feder- ated learning (FL) for cybersecurity threat detection. While FL enables collaborative model training across decentralized nodes without raw data sharing, its reliance on synchronized aggregation introduces latency and vulnerability to stragglers. The AFL framework eliminates synchronization barriers, allowing nodes to contribute updates independently, thereby accelerating convergence and optimizing resource utilization. A comparative evaluation using a standardized cybersecurity dataset (covering malware, phishing, and intrusion attacks) demonstrates that AFL achieves comparable or superior detection accuracy (up to 12% improvement in heterogeneous data scenarios) while reducing convergence time by 30% compared to FL. The asynchronous design also enhances resilience against malicious participants by decentralizing aggregation and minimizing single points of failure. This approach highlights AFL’s scalability and efficiency in real-world applications, offering a privacy-preserving, adaptive solution for dynamic cyber threats.