Efficient Data Erasure in Deep Learning via Sparsity and Stability Mechanisms
Urbana Jaman Orthee, Abu Saleh Muhammad Naeem, Tamanna Khatun, Meharun Ohona, Nafees Mansoor
Conference Paper. 2025 28th International Conference on Computer and Information Technology, ICCIT 2025, pp. 5140–5144 (2025).
Abstract
Machine unlearning (MU) has emerged as a critical paradigm to ensure trained models can selectively remove the influence of specific data without requiring full retraining. While early MU methods addressed privacy regulations such as the 'right to be forgotten', they often suffer from inefficiency, instability, or degraded model performance at scale. To advance MU toward practical deployment, this study introduces a sparsityenhanced unlearning framework that integrates pruning-based model reduction with stability-focused optimization strategies. Two novel methods are proposed: GA Repair, an extension of Gradient Ascent that alternates destructive forgetting with robust repair phases, and FT Confusion, a Fine-Tuning variant that employs entropy maximization and anchored preservation to achieve class-level erasure. Both approaches are evaluated on ResNet-18 with the CIFAR-10 benchmark under class-wise and instance-level forgetting scenarios. Results show that sparsity consistently narrows the gap between approximate unlearning and retraining, improving computational efficiency while maintaining stability. FT Confusion achieves retrain-level forgetting in class-wise tasks, whereas GA Repair provides stronger fidelity in random instance-level forgetting. These findings highlight that no single unlearning strategy is universally optimal; instead, scalable unlearning requires a portfolio of methods, with sparsity as a key enabler for efficiency and robustness. © 2025 IEEE.
Keywords
Deep Learning, Machine Unlearning, Model Sparsification, Optimization