Holistic Unlearning Benchmark: A Multi-Faceted Evaluation for Text-to-Image Diffusion Model Unlearning

As text-to-image diffusion models become advanced enough for commercial applications, there is also increasing concern about their potential for malicious and harmful use. Model unlearning has been proposed to mitigate the concerns by removing undesired and potentially harmful information from the pre-trained model. So far, the success of unlearning is mainly measured by whether the unlearned model can generate a target concept while maintaining image quality. However, unlearning is typically tested under limited scenarios, and the side effects of unlearning have barely been studied in the current literature. In this work, we thoroughly analyze unlearning under various scenarios with five key aspects. Our investigation reveals that every method has side effects or limitations, especially in more complex and realistic situations. By releasing our comprehensive evaluation framework with the source codes and artifacts, we hope to inspire further research in this area, leading to more reliable and effective unlearning methods.

Feature Unlearning for Generative Models via Implicit Feedback

We tackle the problem of feature unlearning from a pretrained image generative model. Unlike a common unlearning task where an unlearning target is a subset of the training set, we aim to unlearn a specific feature, such as hairstyle from facial images, from the pretrained generative models. As the target feature is only presented in a local region of an image, unlearning the entire image from the pretrained model may result in losing other details in the remaining region of the image. To specify which features to unlearn, we develop an implicit feedback mechanism where a user can select images containing the target feature. From the implicit feedback, we identify a latent representation corresponding to the target feature and then use the representation to unlearn the generative model. Our framework is generalizable for the two well-known families of generative models: GANs and VAEs. Through experiments on MNIST and CelebA datasets, we show that target features are successfully removed while keeping the fidelity of the original models.