Fast and Accurate Class-Level Machine Unlearning using Impair–Repair and Noise-Induced Forgetting Mechanisms

Data privacy has become a critical concern in modern machine learning systems, particularly under regulations such as the General Data Protection Regulation (GDPR), which grants individuals the right to erase their data. However, removing specific learned information from trained neural networks without full retraining remains a significant challenge. Existing machine unlearning techniques often struggle to efficiently and completely eliminate targeted information while maintaining overall model performance. In this work, we propose two efficient class-level machine unlearning strategies: Unlearning by Single Impair and (a Few) Repair (USI(F)R) and Direct Noise Forget (DNF). Framed within the cognitive paradigm of ’selective forgetting,’ these methods allow artificial systems to prune specific memory traces while maintaining the integrity of retained knowledge. The first method weakens class-specific representations through a structured impair–repair process, while the second accelerates forgetting by directly injecting normally distributed noise to disrupt targeted model parameters. Both approaches aim to selectively remove class knowledge while maintaining performance on retained classes. We evaluate these methods on four deep learning models across ten benchmark datasets under single-class and multi-class unlearning settings. Experimental results show that the target-class accuracy is reduced to 0% while preserving overall model performance. Additionally, the direct noise method significantly reduces unlearning time from 92 seconds to just 1.5 seconds, making it highly efficient for real-world applications. These findings demonstrate that targeted forgetting mechanisms are effective, scalable, and essential for the development of privacy-compliant, autonomous cognitive systems.