Imperfect Restoration Improves Data Availability Attacks

New research proposes a poisoning method to counter Data Availability Attacks (DAAs) that seeks to improve upon the Convolution-based Unlearnable Dataset (CUDA) method. The study analyzes how sub-optimal gradients in existing methods create class-wise bias that makes data unlearnable.