Targeted Unlearning Using Perturbed Sign Gradient Methods With Applications On Medical Images

Machine unlearning aims to remove the influence of specific training samples from a trained model without full retraining. While prior work has largely focused on privacy-motivated settings, we recast unlearning as a general-purpose tool for post-deployment model revision. Specifically, we focus on utilizing unlearning in clinical contexts where data shifts, device deprecation, and policy changes are common. To this end, we propose a bilevel optimization formulation of boundary-based unlearning that can be solved using iterative algorithms. We provide convergence guarantees when first-order algorithms are used to unlearn. Our method introduces tunable loss design for controlling the forgetting-retention tradeoff and supports novel model composition strategies that merge the strengths of distinct unlearning runs. Across benchmark and real-world clinical imaging datasets, our approach outperforms baselines on both forgetting and retention metrics, including scenarios involving imaging devices and anatomical outliers. This work establishes machine unlearning as a modular, practical alternative to retraining for real-world model maintenance in clinical applications.

💡 Research Summary

**
Machine unlearning has traditionally been explored as a privacy‑preserving tool, often evaluated on small benchmark datasets such as CIFAR‑10 or MNIST. This paper reconceptualizes unlearning as a general‑purpose mechanism for post‑deployment model maintenance, with a particular focus on clinical imaging where data distributions shift, devices become obsolete, and regulatory policies evolve. The authors propose a boundary‑based, targeted unlearning framework that is formulated as a bilevel optimization problem.

In the inner (lower‑level) problem, for each sample (x_i) in the forget set (F), the algorithm searches for a minimal perturbation (\delta) that pushes the point across the decision boundary of the original model (f_{w^0}). Rather than using a single deterministic FGSM step, the authors introduce a perturbed sign‑gradient update:

\