keyword
transferable unlearnable examples
Transferable unlearnable examples are data samples modified with subtle, imperceptible perturbations designed to prevent unauthorized machine learning models from extracting useful patterns, while maintaining this protective capability across diverse model architectures, training hyperparameters, and datasets. In machine learning data privacy, unlearnable examples introduce protective noise that acts as a shortcut during training, causing models trained on the data to fail to generalize to clean test data. Transferable unlearnable examples specifically address the limitation of conventional unlearnable data, whose protective effects often degrade outside a specific target training pipeline, by ensuring that the inability to learn generalizable features successfully transfers to arbitrary, unseen learning algorithms and downstream tasks.
1 item
