DECAF: De-Clustering for Adaptive Representational Unlearning
2026-07-27 • Machine Learning
Machine Learning
AI summaryⓘ
The authors focus on making trained machine learning models forget specific data points completely, which is important for privacy and ethics. They find that existing methods can be tricked by a clustering attack that reveals forgotten data groups. To fix this, the authors introduce DECAF, a method that adds noise and changes output patterns to break these clusters without needing all training data. Testing on an image dataset, DECAF better removes traces of forgotten data while keeping model accuracy high and works faster than other methods.
machine unlearningclustering attackprivacyfeature spacepost-hoc methodCIFAR-10ResNet-18entropyconfidence suppression
Authors
Anjie Le, Can Peng, Hongcheng Guo, J. Alison Noble
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.