Unified Framework for Enhancing Federated Learning Security and Robustness using GANs, Blockchain, and Differential Privacy

Abstract

Neural Network design by a group of clients (each is a person who owns the data) where the data remains private. FL is vulnerable to adversarial attacks, data poisoning, and Byzantine faults, which are threats destroying the integrity as well as the security of the trained model. In order to deal with the challenges above, we propose a new framework, namely FL-GAN-TrustDP, for security enhanced FL using Generative Adversarial Networks (GANs) for adversarial defence, blockchain based hierarchical trust evaluation and adaptive differential privacy. Optimizing the privacy-utility tradeoff based on client trust scores, the adaptive privacy mechanism is a mechanism. GAN based adversarial filtering helps in detecting adversarial updates and thus preventing it, and the trust mechanism backed by blockchain dynamically penalizes the malicious clients. Experimental results show that FL-GAN-TrustDP significantly outperforms baseline FL models (in terms of higher model accuracy, lower adversarial success rates, lower false alarms and faster convergence speed) compared to FedAvg, FedSGD, FedDP, FedBlockchain. In particular, the adversarial success rate on adversarial data seems to decrease significantly from 60% to below 30%, and the notice is more precise, recall, and F1 score than previous work for safeguarding FL. This proposed framework promotes security, privacy, and robustness for FL applications and thus can be a secure federated learning solution in IoT and edge computing environments.