"A Robust Privacy-Preserving Federated Learning Model Against Model Poisoning Attacks" presented by Dr Abbas Yazdinejad, based on IEEE Xplore® article under the same title. Published IEEE Transactions on Information Forensics and Security, June 2024. Visit the article here.
Federated learning enables collaborative model training without direct data sharing, offering an important layer of privacy for distributed machine learning systems. However, these systems remain vulnerable to model poisoning attacks, where malicious participants submit manipulated updates to degrade overall performance. These risks become significantly more severe in privacy-preserving settings; encrypted gradients limit the applicability of traditional defense mechanisms and complicate anomaly detection, particularly under heterogeneous data distributions.
To address these vulnerabilities, this talk presents a robust, privacy-preserving federated learning framework designed to defend against model poisoning attacks while maintaining high efficiency and model accuracy. The proposed approach introduces an internal auditing mechanism capable of analyzing encrypted gradients without exposing sensitive information. By combining additive homomorphic encryption with statistical modeling techniques, the system successfully distinguishes between benign and malicious updates, even in non-independent and non-identically distributed (non-IID) environments.
Furthermore, the framework incorporates probabilistic clustering and distance-based analysis to support Byzantine-tolerant aggregation while minimizing computational and communication overhead. Experimental results across multiple datasets demonstrate that this method significantly improves robustness against both targeted and untargeted poisoning attacks, achieving strong accuracy, scalability, and privacy guarantees. Ultimately, this work highlights how carefully designed encrypted analytics can strengthen the security of federated learning systems without sacrificing practicality.
