Channel-Level Semantic Perturbation: Unlearnable Examples for Multiple Training Paradigms

A study systematically examines the performance of unlearnable examples under multiple training paradigms, and finds that loading and freezing pretrained weights significantly weakens the effectiveness of existing methods under the pretrain-finetune (PF) paradigm. The authors explain this phenomenon by semantic filtering, and propose Shallow Semantic Camouflage (SSC), which restricts perturbation generation to a semantically valid subspace to bypass the semantic suppression of pretrained weights; under shallow freezing and the SF-Pretrain setting, SSC still maintains the unlearnability of the data.