Machine Unlearning Comparator

Machine unlearning aims to remove the influence of specific training data without retraining a model from scratch. In practice, that goal is hard to evaluate: a method can forget targeted samples, damage retain-set utility, leak membership signal under attack, or simply cost too much to run. Comparing methods therefore means reading tradeoffs across accuracy, efficiency, and privacy rather than ranking a single score.

Machine Unlearning Comparator is a web-based visual analytics system designed for that comparative workflow. Researchers move through four stages—Build, Screen, Contrast, and Attack—to configure experiments, scan high-level outcomes, inspect model behavior in detail, and audit residual privacy risk. Linked views connect class-level metrics, prediction matrices, embedding spaces, and layer-wise similarity so analysts can jump from overall method performance to individual samples and representation changes.

The system ships with common unlearning baselines such as Fine-Tuning, Gradient Ascent, Random Labeling, SCRUB, and SalUn, and supports custom methods through a Python registration hook. Interactive Membership Inference Attack views help verify whether forgotten data still leaves detectable traces. The project was developed at the Intelligent Data Computing Lab at SKKU and published in IEEE TVCG 2026.