keyword
perturbative availability poisons
Perturbative availability poisons are subtle, imperceptible modifications added to training data to prevent machine learning models from learning generalizable patterns, effectively rendering the dataset unlearnable. By introducing crafted noise or shortcut features into samples without altering their true labels or human-perceptible content, these perturbations cause learning algorithms to rely on artificial signals rather than meaningful semantic information. As a result, models trained on the poisoned data experience severe performance degradation and fail to generalize when evaluated on clean, unaltered test data. Unlike backdoor attacks that seek to insert targeted vulnerabilities or triggers, perturbative availability poisons aim to deny the overall utility of datasets for model training, commonly functioning as a protective countermeasure to safeguard private data and intellectual property against unauthorized machine learning exploitation.
1 item
