[2509.13007] ReTrack: Data Unlearning in Diffusion Models through Redirecting the Denoising Trajectory

Computer Science > Machine Learning arXiv:2509.13007 (cs) [Submitted on 16 Sep 2025 (v1), last revised 29 Mar 2026 (this version, v2)] Title:ReTrack: Data Unlearning in Diffusion Models through Redirecting the Denoising Trajectory Authors:Qitan Shi, Cheng Jin, Jiawei Zhang, Yuantao Gu View a PDF of the paper titled ReTrack: Data Unlearning in Diffusion Models through Redirecting the Denoising Trajectory, by Qitan Shi and 3 other authors View PDF HTML (experimental) Abstract:Diffusion models excel at generating high-quality, diverse images but suffer from training data memorization, raising critical privacy and safety concerns. Data unlearning has emerged to mitigate this issue by removing the influence of specific data without retraining from scratch. We propose ReTrack, a fast and effective data unlearning method for diffusion models. ReTrack employs importance sampling to construct a more efficient fine-tuning loss, which we approximate by retaining only dominant terms. This yields an interpretable objective that redirects denoising trajectories toward the $k$-nearest neighbors, enabling efficient unlearning while preserving generative quality. Experiments on MNIST T-Shirt, CelebA-HQ, CIFAR-10, and Stable Diffusion show that ReTrack achieves state-of-the-art performance, striking the best trade-off between unlearning strength and generation quality preservation. Comments: Subjects: Machine Learning (cs.LG) Cite as: arXiv:2509.13007 [cs.LG]   (or arXiv:2509.13007v2 [cs.LG...