LoReUn: Data Itself Implicitly Provides Cues to Improve Machine Unlearning

Recent generative models face significant risks of producing harmful content, which has highlighted machine unlearning (MU) as a crucial method for removing the influence of undesired data. However, different difficulty levels among data points can affect unlearning performance. In this paper, we propose that the loss of a data point implicitly reflects its varying difficulty level, leading to our plug-and-play strategy, Loss-based Reweghting Unlearning (LoReUn), which dynamically reweight data throughout the unlearning process with minimal computational effort. Our method significantly reduces the performance gap with exact unlearning in both image classification and generation tasks, effectively enhancing the prevention of harmful content generation from text-to-image diffusion models.