Model Extraction for Image Denoising Networks

Abstract

Model Extraction (ME) replicates the performance of another entity’s pretrained model without authorization. While extensively studied in image classification, object detection, and other tasks, ME for image restoration has been scarcely studied despite its broad applications. This paper presents a novel ME framework for image denoising networks, a fundamental one in image restoration. The framework tackles unique challenges like the black-box nature of the victim model, limiting access to its parameters, gradients, and outputs, and the difficulty of acquiring data that matches the original noise distribution while having adequate diversity. Our solution involves simulating the victim’s noise conditions to transform clean images into noisy ones and introducing loss functions to optimize the generator and substitute model. Experiments show that our method closely approximates the victim model’s performance and improves generalization in some scenarios. To the best of our knowledge, this work is the first to address ME in the field of image restoration, paving the way for future research in this area.