FMU: Fair Machine Unlearning via Distribution Correction
Abstract: Machine unlearning, a technique used to remove the influence of specific data points from a trained model, is often applied in high-stakes scenarios. While most current machine unlearning methods aim to maintain the performance of the model after removing requested data traces, they may inadvertently introduce biases during the unlearning process. This raises the question: Does machine unlearning actually introduce bias? To address this question, we evaluate the fairness of model predictions before and after applying existing machine unlearning approaches. Interestingly, our findings reveal that the model after unlearning can exhibit a greater bias. To mitigate the bias induced by unlearning, we developed a novel framework, Fair Machine Unlearning (FMU), which ensures group fairness during the unlearning process. Specifically, for privacy preservation, FMU first withdraws the model updates of the batches containing the unlearning requests. For debiasing, it then deletes the model updates of sampled batches that have reversed sensitive attributes associated with the unlearning requests. To validate the effectiveness of FMU, we compare it with standard machine unlearning baselines and one existing fair machine unlearning approach. FMU demonstrates superior fairness in predictions while maintaining privacy and comparable prediction accuracy to retraining the model. Furthermore, we illustrate the advantages of FMU in scenarios involving diverse unlearning requests, encompassing various data distributions of the original dataset. Our framework is orthogonal to specific machine unlearning approaches and debiasing techniques, making it flexible for various applications. This work represents a pioneering effort, serving as a foundation for more advanced techniques in fair machine unlearning.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Cedric_Archambeau1
Submission Number: 2784
Loading