Towards Copyright-preserving Dataset Sharing via Dataset Ownership Verification

High-quality open-sourced and commercial datasets significantly prompt AI prosperity. However, existing classical data protection methods (e.g., encryption) cannot protect their copyright by preventing unauthorized use for commercial purposes. In this Webinar, we will introduce the concept of dataset ownership verification (DOV), which is the first and only feasible solution to protect them. In general, DOV consists of two main stages, including dataset watermarking and ownership verification. In the first stage, dataset owners will introduce some imperceptible watermarked samples to generate the released watermarked version of the original dataset, so that all models trained on it will have specific distinctive prediction behaviors on particular samples while having normal behaviors on standard testing samples. In the second stage, given the API of a suspicious third-party deployed model, the dataset owners will detect whether it is trained on the protected dataset by examining its prediction behaviors on verification samples. In this Webinar, we will first introduce the concept of DOV and the basic requirements of dataset watermarks used for DOV. After that, we will illustrate our method designs and their theoretical supports. Next, we introduce INDigo, a novel INN-guided probabilistic diffusion algorithm for arbitrary image restoration tasks. INDigo combines the perfect reconstruction property of INNs with the strong generative capabilities of pre-trained diffusion models. Specifically, we leverage the invertibility of the networks to condition the diffusion process and in this way we generate high quality restored images consistent with the measurements. Finally, to further showcase the capabilities of INNs, we present additional applications where INNs have had a significant impact, including image reflection removal and image steganography.