Mitigating GAN-based poisoning attacks in federated learning using a decentralized multi-layer security framework with blockchain

Abstract

Federated learning (FL) has gained prominence in decentralized artificial intelligence applications; however, its susceptibility to adversarial attacks, particularly GAN-based poisoning, remains a critical challenge. This study proposes a decentralized, multi-layered security framework integrating blockchain-backed validation, anomaly detection, and consensus-driven aggregation to enhance FL security. The framework effectively filters adversarial model updates, leveraging Mahalanobis distance-based anomaly detection and reputation-weighted selection to improve resilience. Experimental results demonstrate superior performance over existing methods, achieving a 98.7% accuracy, a 97.8% detection rate, and a reduced false positive rate of 2.1%. The system maintains efficiency with minimal computational overhead, ensuring practical applicability. The integration of adaptive defenses enhances robustness towards evolving threats, making this framework a comprehensive solution for securing FL in dynamic environments. Future research will focus on optimizing computational efficiency and incorporating privacy-preserving techniques to further strengthen adversarial resilience.