Project ALVI: Privacy-Driven Defences: Federated Learning Security and Authentication

Federated Learning (FL) is a revolutionary approach to decentralized machine learning, enabling collaborative model training across multiple nodes without sharing raw data, thereby enhancing privacy. However, FL systems face significant security challenges, including man-in-the-middle (MITM) attacks, backdoor injections, label inference attacks, and data/model poisoning. To address these threats, we present Project ALVI: Privacy-Driven Defenses, a comprehensive security framework designed to fortify FL systems. This paper integrates four key components: (1) CodeNexa, a dynamic metric verification mechanism that mitigates MITM attacks by ensuring the integrity of model updates; (2) HydraGuard, a hybrid defense mechanism combining trigger inversion and selective layer finetuning to counter backdoor attacks; (3) Confusional Autoencoders with noise addition and Secure Multi-Party Computation (SMC) to protect against label inference attacks in Vertical Federated Learning (VFL); and (4) SECUNID, an anomaly detection framework that leverages LayerCAM and Autoencoders to identify and neutralize data and model poisoning attacks. Experimental evaluations on the MNIST dataset demonstrate the effectiveness of Project ALVI in maintaining model accuracy while significantly reducing the impact of adversarial threats. The integration of these components into a unified framework offers a robust, scalable, and privacy-driven solution for securing FL systems.