Songtao Jiang, Yuan Wang, Ruizhe Chen, Yan Zhang, Ruilin Luo, Bohan Lei, Yeying Jin, Sibo Song, ZhiBo Yang, Jimeng Sun, Jian Wu, Zuozhu Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Malicious Deepfakes pose serious security risks by producing highly realistic forged faces. While numerous countermeasures have been developed to train binary Deepfake classifiers, their limited generalization capacity restricts practical deployment. To proactively defend against Deepfakes, we propose SVS-WM, a Self-Verifiable Semantic Watermarking strategy. The core idea behind SVS-WM is to embed pairs of correlated watermarks within facial semantics, leveraging the inherent fragility of these features, i.e., any semantic modification will disrupt the watermark correlation, thereby enabling robust Deepfake detection. SVS-WM employs a facial semantic disentanglement and reconstruction network, allowing semi-fragile watermarks to be embedded concurrently across multiple semantic levels, including identity and multi-levels of attributes. Specifically, pairs of pseudo-random noise watermarks are adaptively injected into facial attribute and identity features. During propagation stage, the protected image may encounter identity or facial attributes manipulations, we then detect Deepfakes by verifying the correlation result between the decoded attribute watermark and the extracted identity vector. This unique cross-verification mechanism enables authentication without requiring original reference watermark, thereby realizing blind Deepfake detection. Extensive experiments validate the effectiveness of our approach, achieving an average detection accuracy of 98.19% across diverse Deepfake manipulations.

Optimal transport (OT) is an effective technique for mapping between two distributions. Introducing a definition of distance between the distributions, OT provides the best calculable mapping plan, especially from an irregular distribution to a regular one. For this reason, OT has found wide applications, including deep learning, to mitigate some inherent risks in existing techniques. For example, generative adversarial networks (GANs) are widely used to generate medical images but sometimes experience mode collapse, a phenomenon that can be avoided by using OT. We use OT to map latent space distribution in a GAN model to a reasonable Gaussian distribution and generate more convincing and complete medical images to help train medical AI models. Our method consists of three steps. First, we train an auto-encoder as a basic structure of our model. Second, we introduce OT to the trained auto-encoder, which aligns the latent space distribution with a Gaussian distribution. This mapping helps to enforce a more realistic distribution of the generated images, enhancing their clinical relevance. Third, we construct a GAN-type model by training the decoder part of the auto-encoder and discriminator networks. We conducted a comprehensive experimental evaluation on a diverse medical dataset by comparing the performance of our proposed method to other state-of-the-art models. Results show that our method performs relatively better than other models by FID scores and provides more detailed information.

Recently, there has been a significant surge of interest in medical image generation. In this study, we developed a model known as AE-COT-GAN (autoencoder-based conditional optimal transport generative adversarial network) to generate medical images that belong to specific categories. The primary objective of our research is to address the prevalent challenges often encountered during the training of generative adversarial networks (GANs), including issues such as mode collapse and mode mixing. The training process of our model encompasses three fundamental components. First, we employ an autoencoder model to obtain a low-dimensional manifold representation of real images. Second, we apply extended semi-discrete optimal transport to map Gaussian noise distribution to the latent space distribution and obtain corresponding labels effectively. This procedure leads to the generation of new latent codes with known labels. Finally, we integrate a GAN to train the decoder further to generate medical images. To evaluate the performance of the AE-COT-GAN model, we conducted experiments on two medical image datasets, namely DermaMNIST and BloodMNIST. The model’s performance was compared with state-of-the-art generative models. Results show that the AE-COT-GAN model had excellent performance in generating medical images. Moreover, it effectively addressed the common issues associated with traditional GANs.