DECAF: De-Clustering for Adaptive Representational Unlearning

2026-07-26

AI Summary

This work addresses the vulnerability of existing machine unlearning methods to clustering attacks in continual deployment settings, where reliably removing the influence of specific data remains challenging. The authors propose a post-processing unlearning mechanism that relies solely on the forget set and is the first designed explicitly to counter clustering attacks. Without requiring model retraining or access to the full training dataset, the method disrupts class-cluster structures in feature space through input noise injection, confidence suppression, and entropy-based output diversification, thereby achieving effective representation disentanglement. Evaluated on CIFAR-10, the approach reduces accuracy on the forgotten class to 0.10% while maintaining 79.4% accuracy on retained classes, achieving an Area Under the Stability curve (AUS) of 0.88—significantly outperforming current state-of-the-art methods.

Abstract

Machine unlearning, which aims to remove the influence of specific training data from a trained model, is a key requirement for privacy, accountability, and adaptive deployment. We argue that many unlearning methods are vulnerable to a simple clustering attack, which can recover class structure in an unsupervised manner, limiting their suitability for continual deployment where removal requests must be handled reliably on demand. To address this, we propose DECAF (DE-Clustering for Adaptive Forgetting), a post-hoc method that operates only on the forget set and is designed to break the cluster. DECAF combines input noise, confidence suppression, and entropy-based output diversification to disrupt the residual feature-space structure associated with forgotten data. On CIFAR-10 with ResNet-18, DECAF attains 0.10% forget-class accuracy, 79.4% retain accuracy, and an AUS of 0.88, surpassing all other baselines. In cluster-based analysis, it attains performance comparable to that of unlearning methods that use the full training set, while being significantly more efficient. Code: https://github.com/ale256/representation_unlearning.

Problem

• machine unlearning
• clustering attack
• data removal
• privacy
• adaptive deployment

Innovation

• machine unlearning
• de-clustering
• representation forgetting
• privacy-preserving learning
• post-hoc unlearning