Forgetting at the edge of chaos: Chaotic perturbations for selective machine unlearning

Berta, M., Vargas-Solar, G., & Cerquitelli, T.
PKDD/ECML Workshops
2026

Machine unlearning has emerged as a key requirement for modern machine learning systems, driven by privacy regulations and the growing need to remove specific training data from deployed models. Current methods attempt to forget information while preserving the utility of the remaining knowledge. Despite recent progress, effective forgetting in deep neural networks remains difficult. Existing approaches often rely on costly retraining, or on perturbation-based methods that, while more efficient, can erase targeted information at the cost of damaging retained representations and degrading model behavior.

As a result, forgetting is still poorly understood as a dynamical process in neural networks. In this work, we explore chaos as a mechanism for selective machine unlearning. We introduce a chaotic impairment–repair framework in which deterministic chaotic perturbations destabilize forget-related representations, followed by repair dynamics that recover retained functionality. Across multiple benchmarks, chaotic perturbations often preserve utility more effectively and, in several settings, remain significantly closer to retrained references than matched Gaussian baselines.

[Berta et al., 2026] Berta, M., Vargas-Solar, G., & Cerquitelli, T. (in press). Forgetting at the edge of chaos: Chaotic perturbations for selective machine unlearning. PKDD/ECML Workshops 2026, Naples, Italy, September 7–11, 2026. Accepted for publication.