CAREER: Machine Unlearning in Diffusion Process: Towards a Systematic Design for Quality, Efficiency and Scalability

Generative artificial intelligence has rapidly transformed how people create images, designs, educational materials, and personalized digital services. Diffusion models, as one of the most powerful families of generative AI models, can produce highly realistic and diverse content by learning from large-scale data. However, once these models are trained, different types of learned knowledge often become deeply interconnected within the model and its generation process. As a result, it is difficult to selectively remove sensitive, outdated, private, copyrighted, harmful, or otherwise unwanted concepts without retraining the model from scratch or damaging its overall generation quality. This limitation creates important challenges for privacy protection, responsible AI deployment, user control, and long-term model maintenance. This project develops a systematic machine unlearning framework for diffusion-based generative AI, enabling models to forget specified concepts while preserving their ability to generate useful and high-quality content. Together, these advances will provide a foundation for more trustworthy, controllable, and adaptable generative AI systems. The project will also support education and workforce development by integrating responsible generative AI topics into undergraduate and graduate curricula, engaging K-12 and undergraduate students through accessible research experiences, and broadening participation in trustworthy AI research. This project advances the science of machine unlearning in diffusion models through an integrated research plan that addresses quality, efficiency, scalability, and unified generation control. First, the project will develop high-quality multi-concept unlearning methods based on dynamic gradient masking and concept-aware loss functions. These methods aim to identify and update concept-relevant neurons while preserving unrelated knowledge, and to guide post-unlearning generation toward semantically meaningful alternatives rather than noisy or irrelevant outputs. Second, the project will develop an inference-time unlearning framework to avoid the latency and cost of sequential fine-tuning. It uses neuron saliency maps to build concept-specific masks, fuses these masks to remove multiple concepts without additional training, and applies soft masking, residual low-rank refinement, and entropy-based retention to preserve generation quality and semantic coherence at larger scale unlearning operations. Third, the project will explore a unified design space for concept unlearning and generation control by studying the roles of self-attention and cross-attention in diffusion models and developing content-aware masking and plug-and-play inference-time editing modules for frozen or proprietary models. The resulting framework will be evaluated across multiple diffusion backbones and generative tasks, including text-to-image synthesis, style transfer, image inpainting, and super-resolution. This research will contribute new algorithms, system-level insights, open-source tools, and educational materials for building responsible and controllable generative AI systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.