Channel-Level Semantic Perturbations: Unlearnable Examples for Diverse Training Paradigms
Researchers have introduced a new method called Shallow Semantic Camouflage (SSC) to create unlearnable examples that can resist model training, even when using pre-trained models. Existing unlearnable example techniques are less effective when models are pretrained and then fine-tuned, as the frozen layers preserve semantics and filter out the noise. SSC aims to bypass this by generating perturbations within a semantically valid subspace, ensuring data remains unlearnable across various training paradigms. AI
IMPACT Introduces a new technique to enhance data privacy in AI training, potentially impacting how datasets are secured against unauthorized use.