潛在擴散模型中的內嵌式浮水印方法:使用者歸因與篡改定位應用

Abstract

隨著近年生成式人工智慧的快速發展，利用公開應用程式介面快速產生高品質圖像已成為常態。然而，這樣的便利也伴隨個資洩漏與圖像遭竄改等安全風險。雖然傳統的圖像浮水印技術能達到使用者身分識別與篡改區域標記，但這類「後處理式」方法需在每張圖像生成後再額外嵌入資訊，不僅操作繁瑣，效率也非常低落，不利於大規模應用。 為解決上述問題，本文提出一種全新的「生成中嵌入式」方法，直接在圖像產生過程中，將使用者身分與篡改標記模板一併嵌入，無需額外後處理，大幅提升整體效率。我們的做法是針對變分自動編碼器中的解碼器進行微調，使其生成圖像時自動帶有不可見的識別與定位資訊。同時，我們引入模擬失真層，透過裁切、縮放、壓縮、亮度調整等常見圖像變化來模擬真實環境中可能出現的破壞行為，提升整體的穩定性與耐用性。 實驗結果顯示，本方法能在維持生成圖像視覺品質的前提下，能準確還原使用者身分並精確標示被竄改區域，同時展現出更高的嵌入效率與執行效能，證實其在實務應用上的潛力與價值。

With the rapid advancement of generative artificial intelligence, generating high-quality images through public APIs has become increasingly common. However, this convenience also brings serious security concerns like personal privacy leakage and unauthorized image manipulation. Although traditional post-generation watermarking methods can achieve user attribution and tamper localization, they require additional data embedding for each generated image, resulting in low efficiency and limited scalability in large-scale applications. To address this issue, this work presents a novel in-generation watermarking framework that embeds invisible user-specific identification and localization templates directly during the image generation stage. Specifically, we fine-tune the decoder of a Variational Autoencoder within a Latent Diffusion Model, enabling the generated images to carry the embedded invisible information without requiring any post-processing. This design significantly improves the integration capability and operational efficiency of watermarking. To enhance the system's robustness, we introduce a noise simulation layer during training, which emulates common image manipulations such as cropping, resizing, JPEG compression, and brightness adjustment. This extra layer enables the model to remain effective under divergent real-world scenarios. Experimental results demonstrate that our method maintains high perceptual quality of the generated images while achieving accurate and robust user attribution and tamper localization simultaneously. Moreover, our in-generation strategy outperforms existing benchmarked works simultaneously in terms of performance and efficiency.