FAST: Feature Aware Similarity Thresholding for Weak Unlearning in Black-Box Generative Models

The heightened emphasis on the regulation of deep generative models, propelled by escalating concerns pertaining to privacy and compliance with regulatory frameworks, underscores the imperative need for precise control mechanisms over these models. This urgency is particularly underscored by instances in which generative models generate outputs that encompass objectionable, offensive, or potentially injurious content. In response, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">machine unlearning</i> has emerged to selectively forget specific knowledge or remove the influence of undesirable data subsets from pretrained models. However, modern <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">machine unlearning</i> approaches typically assume access to model parameters and architectural details during unlearning, which is not always feasible. In multitude of downstream tasks, these models function as black-box systems, with inaccessible pretrained parameters, architectures, and training data. In such scenarios, the possibility of filtering undesired outputs becomes a practical alternative. Our proposed method <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">feature aware similarity thresholding (FAST)</i> effectively suppresses undesired outputs by systematically encoding the representation of unwanted features in the latent space. We employ user-marked positive and negative samples to guide this process, leveraging the latent space's inherent capacity to capture these undesired representations. During inference, we use this identified representation in the latent space to compute projection similarity metrics with newly sampled latent vectors. Subsequently, we meticulously apply a threshold to exclude undesirable samples from the output.