OPTML-Group/Unlearn-Saliency

Gradient-based weight saliency identifies and modifies only the most influential model parameters for unlearning, enabling efficient forgetting of data points, classes, or concepts without full retraining. The approach unifies classification and generative models—demonstrated on image classifiers, DDPM with classifier-free guidance, and Stable Diffusion—achieving near-exact unlearning performance (0.2% gap on CIFAR-10) while nearly eliminating harmful image generation in diffusion models. Implementation includes modular task pipelines for both discriminative and generative domains with PyTorch backends.

143 stars.