Self-Erasing Neural Networks (SENNs): A Neurogenesis-Inspired Framework for GDPR-Compliant Machine Unlearning

Data privacy laws and the growing computational burden of large neural networks demand efficient machine unlearning. We propose Self-Erasing Neural Networks (SENNs), a novel biologically-inspired framework employing Neurogenesis-inspired Synaptic Pruning and Dynamic Regrowth (NSP-DR). Our comprehensive study uncovers a critical unlearning trade-off: SENNs achieve unparalleled erasure effectiveness, demon-strated by the highest perplexity on forgotten data, while sig-nificantly boosting efficiency by reducing unlearning time by over 50% and computational resource usage compared to full retraining. Moreover, SENNs robustly reduce privacy leakage, evidenced by lower Membership Inference Attack success. This aggressive forgetting, however, correlates with a degradation in retained model utility. As foundational research, our work positions SENNs as a powerful and practical tool for high-priority erasure scenarios, establishing a clear frontier and future research directions for optimizing this inherent erasure-utility balance.