Image Restoration for Training Data Reconstructed from Trained Neural Networks
Keywords: image restoration, diffusion, privacy attacks, dataset reconstruction
TL;DR: We demonstrate how diffusion based image restoration can be used to significantly improve the quality of images that correspond to training data which has been reconstructed from a given trained neural network.
Abstract: Haim et al. [NeurIPS 2022] propose a method to reconstruct training data from trained neural networks with impressive results. While their reconstructed images resemble the original training images, most of them also contain a considerable amount of noise and artifacts. This is especially true, when the network was trained on more than just a few dozen images. To address this, we view the problem as a specific image restoration task. Since the noise and artifacts are different from other types of noise (Gaussian noise, compression artifacts, blurring, or impulse noise from digital cameras), we create a new dataset specifically for the restoration of images produced by the reconstruction process proposed by Haim et al. We use this dataset consisting of about 60 million noisy reconstructions of CIFAR-10 images to train a diffusion model on the restoration task. Using this method, we obtain reconstructions that are significantly closer to the original training images measured in terms of SSIM and HaarPSI scores.
Supplementary Material: zip
Primary Area: alignment, fairness, safety, privacy, and societal considerations
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.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 12702
Loading