The advancement of deep learning necessitates stringent data privacy guarantees. Dataset distillation has shown potential in preserving differential privacy while maintaining training efficiency. This study first identifies that data generated by state-of-the-art dataset distillation methods strongly resembles to real data, indicating severe privacy leakage. We define this phenomenon as explicit privacy leakage. We theoretically analyze that although distilled datasets can ensure differential privacy to some extent, a high \IPC can weaken both differential privacy and explicit privacy. Furthermore, we reveal that the primary source of privacy leakage in distilled data stems from the common approach of initializing distilled images as real data. To address this, we propose a plug-and-play module, Kaleidoscopic Transformation (KT), designed to introduce enhanced strong perturbations to the selected real data during the initialization phase. Extensive experiments demonstrate that our method ensures both differential privacy and explicit privacy, while preserving the generalization performance of the distilled data. Our code will be publicly available.
Keywords: Dataset Distillation
Abstract:
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 10266
Loading