WaterLo: Protect Images from Deepfakes Using Localized Semi-Fragile Watermark
Nicolas Beuve, Wassim Hamidouche, Olivier Déforges
Abstract
Most existing contributions in the field of Deepfake detection focus on passive detection methods, where the detector only analyzes the doctored image. However, this approach often lacks the ability to generalize to unseen data and struggles to detect Deepfakes generated using new deepfake models. To address this limitation, our paper proposes an active detection approach, where we have access to the image before the Deepfake is generated. Our solution involves applying a watermark that disappears in modified regions, allowing our detector to identify image modifications and localize them accurately. Additionally, we incorporate a compression module into our training pipeline to enhance the watermark’s robustness against JPEG compression. Experimental results demonstrate the effectiveness of our proposed solution, achieving a remarkable detection accuracy of 97.83% while maintaining significantly higher image quality compared to previous works. Furthermore, by incorporating the compression module in the training pipeline, we improve the detection accuracy on compressed samples, albeit with a slight decrease in accuracy for non-compressed samples. This contribution also provides a valuable tool for video owners to verify if their videos have been tampered with and safeguard them against unauthorized use. The code of the proposed framework is available at https://github.com/beuve/waterlo.