Diagnosing Forgetting in AI Models Without Original Data

The field of machine unlearning aims to remove specific information, like a "forget class," from a model while preserving its general knowledge. However, simply achieving low accuracy on the forget class doesn't guarantee that the model has truly erased its understanding of that class. Prior work has shown that these "forgotten" classes can sometimes be recovered, but typically require access to original data or auxiliary information. This paper investigates class relearning in a strictly source-free environment, meaning no original forget or retain samples, auxiliary data, or reference checkpoints are used. The core question is whether a forget class can be recovered by only updating the classifier head of an unlearned model. The researchers propose a theoretical framework and a practical tool called Source-Free Relearning Audit (SFRA). SFRA generates synthetic embeddings and uses model confidence to construct probes, then updates the classifier. To quantify the extent of forgetting, they introduce the Relearning Score (RS), which measures both forget-class recovery and retain-accuracy preservation. Experiments across various datasets and models demonstrate that several approximate unlearning methods leave substantial recoverable structure, sometimes even exceeding the recoverability of a model that was never trained on the class.