AI-Integrated Federated Learning Architecture for Confidential Cyber security Event Prediction

Abstract

The growing complexity of cyber threats requires the creation of predictive models that are capable of integrating knowledge based on decentralized datasets without information disclosure. The traditional centralized methodology is hindered by strict privacy laws and the cyber security risks that they bring. We answer this question by proposing an AI-based Federated Learning (AIFL) framework which is purpose-built and optimized to ensure the confidential prediction of cyber security incidents, and which is optimally configured as a multi-tier hierarchical structure, synergizing differential privacy, secure multi-party computation (SMPC), and homomorphic encryption. AIFL promotes the joint model training without revealing the local data. As an additional iteration of this process, we weaken Privacy-Aware Weighted Federated Averaging (PAW-Fed Average), which allocates weights to the contribution of clients according to data integrity and the amount of privacy loss measurably caused. Empirical analyses of the CIC-IDS2017 and TON-IOT baseline datasets reveal that AIFL can achieve predictive performance within 2.1% of the centralized variants, enhance data privacy surpassing more than 95%, and also optimize communication overheads, by about 30 percent, as compared to the conventional Fed Avg, and creates a viable framework to realize secure, collaborative cyber-defense.