Adversarial Information Bottleneck Training for Privacy-Preserving Federated Learning: An Expanded Framework and Empirical Analysis

This paper proposes AIBT (Adversarial Information Bottleneck Training), a method for training machine learning models across multiple devices (federated learning) while keeping user data private. The problem it solves: In federated learning, devices (phones, hospitals, etc.) train a shared model without sending raw data to a central server. However, the model updates they do send can still leak sensitive personal information. Existing privacy solutions like Differential Privacy protect privacy but hurt model accuracy significantly. How AIBT works: It combines two techniques: Information Bottleneck — compresses the model's internal representations to keep only what's needed for the task, discarding irrelevant details Adversarial training — adds a component that actively tries to predict sensitive attributes (like gender or race), while the main model is trained to fool it, effectively hiding that information