NeuroSymbolic approaches to machine unlearning: enhancing selective forgetting through hybrid AI systems

Abstract

Machine unlearning has emerged as a critical requirement to satisfy privacy regulations like the GDPR, which mandate that personal data or malicious artifacts be ”forgotten” by trained models. While exact unlearning (retraining from scratch) offers a theoretical guarantee of data removal, it is often computationally infeasible for large models. Conversely, standard approximate unlearning methods, such as Gradient Ascent, typically suffer from catastrophic forgetting, rendering the model useless. We propose Neuro-Symbolic Amnesiac Unlearning (NS-AU), a novel approximate unlearning technique for robust image classification that utilizes a logicguided penalty term to surgically remove targeted data contributions. By integrating symbolic constraints directly into the loss landscape, our system can reason about inconsistent feature associations introduced by poisoning attacks. Experimental results on a VGG architecture demonstrate that NS-AU achieves 82.54% accuracy outperforming even the exact ”Gold Standard” baseline while reducing the Attack Success Rate (ASR) of backdoor triggers from 98.40% to just 3.00%. This hybrid scheme offers a superior balance of efficiency and stability compared to traditional approximate methods, ensuring robust defense against adversarial backdoors.