Reversible Unlearnable Examples: Towards the Copyright Protection in Deep Learning Era

2026-08-06

AI Summary

In the era of deep learning, data copyright faces dual threats from unauthorized model training and malicious data leakage, yet existing approaches struggle to simultaneously achieve effective defense and reliable watermark extraction. This work proposes a unified framework that generates invertible unlearnable examples to concurrently prevent unauthorized training and enable robust copyright watermark embedding and retrieval. The core innovation lies in a perturbation generation mechanism based on mutual information minimization, which forces models to learn irrelevant features and thereby suppresses effective training, combined with a dual-extractor watermarking strategy that ensures perturbation invertibility and high-accuracy copyright verification. Experiments demonstrate that the method significantly enhances copyright protection on ImageNet, CIFAR-10, and Pets datasets while maintaining high watermark extraction accuracy.

Abstract

Significant advancements in deep learning have been made possible by the utilization of large datasets, underscoring the critical importance of copyright protection. Adding meticulously designed perturbations to examples, making them unlearnable has become a crucial approach for safeguarding data copyright. Existing methods for creating unlearnable examples overlook the risk of data leakage, which can threaten data ownership. Thus, copyright protection in deep learning faces two main threats: illegal model training and malicious data leakage. We investigate that these two threats cannot be solved by straightforwardly combining existing availability attacks and watermarking techniques as their negative interaction effects. Therefore, in this paper, we propose a novel copyright protection mechanism for the aforementioned security concerns. Considering that the prevention of unauthorized model training requires powerful generalizability of unlearnable perturbations, we generate perturbations to induce the model to learn uncorrelated features of input images. It works by minimizing the mutual information of the input and output of the model. On the other hand, to eliminate the side impact of unlearnable perturbations on the watermark extraction, we design a dual extraction strategy by using two distinct watermark extractors. Extensive experiments on the image datasets {ImageNet, CIFAR10, and Pets} show that our proposed method could provide comprehensive copyright protection to images. The code is available at {https://github.com/Yeah21/ReversibleUnlearnableExamples}.

Problem

• copyright protection
• unlearnable examples
• data leakage
• deep learning
• model training

Innovation

• Unlearnable Examples
• Copyright Protection
• Mutual Information Minimization
• Dual Watermark Extraction
• Reversible Perturbations