Learnability and Privacy Vulnerability are Entangled in a Few Critical Weights

TL;DR

This paper reveals that only a small subset of neural network weights are both critical for utility and vulnerable to privacy attacks, proposing a targeted rewinding method to enhance privacy without sacrificing performance.

Contribution

It introduces a novel approach focusing on critical weights' locations for privacy preservation, outperforming existing methods in resisting membership inference attacks.

Findings

Selective weight rewinding improves privacy protection.

Method maintains high utility compared to full retraining.

Critical weights are more about location than value.

Abstract

Prior approaches for membership privacy preservation usually update or retrain all weights in neural networks, which is costly and can lead to unnecessary utility loss or even more serious misalignment in predictions between training data and non-training data. In this work, we observed three insights: i) privacy vulnerability exists in a very small fraction of weights; ii) however, most of those weights also critically impact utility performance; iii) the importance of weights stems from their locations rather than their values. According to these insights, to preserve privacy, we score critical weights, and instead of discarding those neurons, we rewind only the weights for fine-tuning. We show that, through extensive experiments, this mechanism exhibits outperforming resilience in most cases against Membership Inference Attacks while maintaining utility.