Divergence-Induced Contrastive Unlearning via Directed Representation Shifts

Jayasree Saha, Xiaoyi Jiang

Keywords: Unlearning, Data Poisoning, indiscriminate attack, contrastive Learning

Abstract: Unauthorized data collection has become widespread, raising the need for defenses that prevent exploitation of personal data. Unlearnable Examples (UEs) address this by embedding imperceptible perturbations that preserve visual quality while making data unusable for training. Recent work has shown that contrastive learning can be poisoned to generate UEs, but existing methods lack theoretical grounding and fail to exploit the geometric structure of learned representations. In this work, we present the first principled analysis of contrastive poisoning and reveal why it is effective. Building on this understanding, we propose Divergence-Induced Contrastive Unlearning (DICU), a framework that introduces direction-aware divergence regularization into the poisoning objective. This design amplifies intra-class sparsity, pushes samples beyond class manifold boundaries, and enables free mixing across classes, producing stealthy and robust perturbations. Our approach is especially effective in high class-count settings, reducing linear probing accuracy at significant level.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 7868