加強潛在擴散模型中根源浮水印的強韌性

Abstract

近期許多研究人員深入探索複雜的潛在擴散模型領域，這些模型能依據多模態條件生成逼真的圖片，這種特性賦予了它們強大的潛力，適用於各種應用。同時，這類模型也牽涉到迫在眉睫的倫理困境，主要受限於如何負責任地部署這項技術。惡意使用者製造的圖像有可能造成重大的社會問題，需要嚴格的控制措施以減輕可能的危害。根源浮水印技術是解決此問題可行的方法，這項技術允許我們追踪那些惡意創建的圖像來源，從而利用法律向為非作歹為非作歹者咎責。本篇論文基於現有的方法加以改善，著重於提升根源浮水印對於圖像失真的抵抗能力。通過對現有方法模組的正規化和在訓練期間加入圖像噪聲，本篇論文加強了根源浮水印的強韌性同時保持生成圖像的視覺品質。研究結果顯示，在圖片縮小為原圖的10%後，浮水印解碼之為元準確率相較現有方法高出24%，此結果展現本篇論文的有效性，也在潛在擴散模型之根源浮水印領域中取得重要的進展。

Recently, many researchers dived into the intricate realm of latent diffusion models, renowned for their competence in generating images reflecting a multimodal range of conditions. This property endows them with enormous potential, rendering them fit for a host of applications across various disciplines. Simultaneously, the power of such models introduces pressing ethical dilemmas, primarily revolving around the responsible deployment of this technology. The indiscriminate use of such models has the potential to cause significant harm, emphasizing the need for rigorous control measures to mitigate potential misuse. Root watermarking is introduced as a countermeasure to address these ethical concerns. This technology allows for the tracing of images that have been maliciously created back to the originators. By establishing this link, we can assign responsibility for any misuse, facilitating the enforcement of legal sanctions and deterring misuse. In this work, we focus on enhancing the resilience of common distortions in the generated images. We have expanded upon an existing method, refining it through the normalization of its module and the strategic injection of noise during the training. This approach has allowed us to significantly enhance the model’s robustness, all the while maintaining the perceptual quality of the generated images. Our quantitative analysis attests to the efficacy of our approach, as it has resulted in a substantial improvement in bitwise accuracy by up to 24%, even when the images undergo a 10% resizing. This outcome not only showcases the effectiveness of our method but also represents a significant stride forward in the ongoing journey to watermarking latent diffusion models.