Learnability and Privacy Vulnerability are Entangled in a Few Critical Weights

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Fang, Xingli, Kim, Jung-Eun
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908884556316672
author Fang, Xingli
Kim, Jung-Eun
author_facet Fang, Xingli
Kim, Jung-Eun
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2603_13186
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Learnability and Privacy Vulnerability are Entangled in a Few Critical Weights
Fang, Xingli
Kim, Jung-Eun
Machine Learning
Artificial Intelligence
Cryptography and Security
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.
title Learnability and Privacy Vulnerability are Entangled in a Few Critical Weights
topic Machine Learning
Artificial Intelligence
Cryptography and Security
url https://arxiv.org/abs/2603.13186