DECAF: De-Clustering for Adaptive Representational Unlearning

Researchers introduced DECAF, a post-hoc machine unlearning method designed to disrupt residual clustering in forgotten data.

It works only on the forget set by adding input noise, suppressing confidence, and diversifying outputs with entropy-based objectives.

On CIFAR-10 with ResNet-18, DECAF achieved very low forget-class accuracy, strong retain accuracy, and competitive resistance to cluster attacks.

The approach addresses a key weakness in current unlearning methods that can still expose class structure after data removal, improving privacy and reliability.