Ethical AI for Unlearning the Biased and Sensitive Information

Abstract

The rapid advancement of artificial intelligence systems across sensitive domains has elevated serious concerns about the retention of biased, private, and ethically problematic information within trained models. This research presents an ethical AI framework for unlearning biased and sensitive information from machine learning models. An initial BERT model was trained to achieve a good reference performance in the classification task. Gradient ascent-based unlearning was used to unlearn a certain amount of targeted information to eliminate it. Findings revealed that unlearning in itself resulted in the apparent reduction of accuracy and the inability to separate classes. In order to regain model stability, retraining was done with fine-tuning on the retained dataset. The overall performance of the fine-tuning was the best, as it enhanced precision and minimized the prediction errors. An intermediate measure to harmonize forgetting and performance recovery was also tested as hybrid method of unlearning and retraining. The effectiveness of retraining-based recovery was proved by comparative analysis with accuracy, precision, recall, F1-score, ROC-AUC, and confusion matrices. The results show that to perform ethical machine unlearning, effective forgetting and performance restoration are needed. This piece of work facilitates building reliable AI systems in accordance with the requirements of fairness and privacy.