Research: Evaluating Secure Aggregation Techniques for Privacy-Preserving Federated Lea...

Research question: How effective are secure aggregation protocols in preserving privacy and model accuracy during federated learning under adversarial threat scenarios?

Federated learning enables collaborative model training across distributed devices without sharing raw data, addressing privacy concerns central to modern machine learning applications. However, the aggregation of model updates can still expose sensitive information, especially when adversaries attempt to infer data from transmitted gradients.

Current research focuses on secure aggregation methods—such as homomorphic encryption and secret sharing—to mitigate privacy leakage, yet their practical effectiveness and trade-offs in real-world adversarial settings remain underexplored. Existing studies often emphasize theoretical guarantees or simplistic threat models, leaving a gap in empirical evaluation under realistic attack scenarios.

This project will conduct a systematic review of secure aggregation protocols, design experiments simulating federated learning with varying adversarial threats, and empirically assess privacy preservation and model fidelity. The study will utilize public datasets, implement state-of-the-art aggregation schemes, and analyze the impact of adversarial attacks on both privacy and accuracy metrics.

Understanding these trade-offs is critical for deploying federated learning in sensitive domains such as healthcare and finance, where privacy breaches can have severe consequences. The project aims to inform protocol selection and future research directions in secure collaborative learning.