Efficient unlearning for data security in deep learning systems
Enting Guo, Chunhua Su, Peng Li · The Computer Journal · 2025
Abstract Machine unlearning in the context of cybersecurity and privacy protection facilitates the removal of specific training data impacts from deep learning (DL) models, adhering to security, privacy, or compliance demands. However, traditional methods can only handle short-term, independent unlearning tasks. Conversely, real-world scenarios often involve extensive unlearning demands from users. Current methods fail to adequately address these demands due to substantial computational overhead and adverse impacts on inference accuracy, leaving the security and privacy of many users at risk. To navigate these challenges adeptly, we introduce the Multi-Agent Reinforcement Learning Data Lifecycle Management (MADLM) strategy. MADLM intricately examines the interactions between unlearning and continuous learning processes, enabling the postponement of certain tasks for combined execution to optimize computational resources. Concurrently, it employs strategic data management to maintain and enhance inference accuracy. Furthermore, by utilizing Multi-Agent Reinforcement Learning (MARL), MADLM dynamically orchestrates task scheduling to minimize computational demands, improve task response times, and bolster inference reliability, crucial for upholding stringent cybersecurity and privacy standards. Our evaluations of MADLM reveal substantial enhancements, including a 6% uplift in inference accuracy and a dramatic reduction in computational overhead to merely 12% of the original demands, effectively expanding the data security protections.