Machine Unlearning: Towards Privacy-Preserving and Trustworthy Artificial Intelligence System

Abstract

Artificial intelligence models today depend heavily on large datasets, where information becomes part of the model during training. Once this information is learned, removing the effect of specific data becomes challenging. This creates concerns related to user privacy, data protection, and legal requirements for data removal. Existing solutions, such as retraining or fine-tuning, are often time-consuming and may not fully remove the data’s impact. To address this issue, this work proposes a Dual-Layer AI Controlled Unlearning System (DSRUN) that allows faster and targeted data removal. The system separates the learning and unlearning processes and uses a data influence tracking mechanism to understand how individual data points affect model parameters. Instead of retraining the entire model, only the affected components are updated, reducing computational effort while maintaining performance. Experimental results show that the proposed system achieves faster unlearning with minimal impact on accuracy and reduces the chances of recovering removed data. This approach provides a practical solution for building scalable and privacy-aware AI systems capable of supporting real-time data removal.