中国科学院自动化研究所机构知识库

       在数字化和多媒体技术的推动下，图像成为重要的信息传递和创意表达载体。先进的图像编辑技术和图像生成技术的发展简化了逼真视觉作品的创作，推动了电影和虚拟现实等行业的进步。然而，这些技术的进步也让伪造图像变得更加容易便捷，模糊了真伪之间的界限。精心制作的伪造图像，特别是伪造人脸，容易误导公众而引发恐慌和误解，甚至被用于政治操纵、欺诈诽谤等恶意目的，对个人和社会安全构成严重威胁。因此，随着技术普及和滥用风险增加，开发有效的伪造图像鉴别技术以验证图像的真实性，成为一项必要且紧迫的重要任务。

       伪造图像鉴别面临多重挑战，需应对不同的伪造形式，包括全图生成的计算机生成图像和部分篡改的局部篡改图像。尽管现有鉴别方法有所成效，但面对日益精进的伪造技术时仍存在局限：（1）在计算机生成图像鉴别中，现有数据集质量不佳且方法侧重于全局特征，忽视关键的细节差异；（2）在伪造人脸鉴别中，现有方法过于关注特定伪造痕迹而忽略不同伪造技术的共性特征；（3）新型伪造人脸数据难以获取，导致模型难以对其进行针对性训练；（4）现有方法未能充分利用面部单元间的关联性，影响了对人脸特性的深入理解。针对以上问题，本文提出对应的创新方法，研究内容归纳如下：

        1. 针对计算机生成图像鉴别中存在的数据集和模型局限性，本文构建了大规模计算机生成图像数据集（Large-Scale Computer-Generated image Benchmark, LSCGB），并提出基于纹理一致性的计算机生成图像鉴别方法。LSCGB数据集具有数据规模大、样本多样性高及类别偏差低的特点，为该领域提供了更具挑战性的测试平台。在此基础上，通过深入分析计算机生成图像与真实图像间的纹理特征差异，提出了基于纹理一致性的鉴别模型，以强化输入特征中的纹理信息，并在多个尺度上整合纹理特征，以精确区分计算机生成图像与真实图像。与现有方法对比以及消融实验结果表明，所提方法在跨数据集的泛化能力以及对图像后处理操作的鲁棒性方面具备有效性。

       2. 面对现有伪造人脸鉴别方法忽视共性特征的问题，本文提出了基于噪声一致性的伪造人脸鉴别方法。现有的人脸伪造过程会在面部与背景之间引入显著的噪声不一致。基于这一发现，本文提出基于噪声一致性的鉴别方法，该方法包含噪声一致性增强和多尺度噪声一致性学习两个核心模块。前者通过增强人脸与背景之间的噪声差异来增强特征表征能力，而后者则在多个尺度上计算局部区域间的噪声一致性，并通过多尺度特征集成策略来整合信息，以提高伪造鉴别的精确度。经过在多个数据集上的广泛测试，所提方法在伪造人脸鉴别的准确性和泛化性方面都取得了性能提升。

       3. 针对新型伪造人脸数据难以获取的问题，本文提出基于退化过程一致性的伪造人脸鉴别方法，减少对伪造人脸训练数据的依赖。图像在采集和传输过程中会经历不同退化过程，因而携带不同退化信息。伪造算法通常将不同来源的人脸与背景进行拼接，导致这些区域存在退化过程的不一致。基于此观察，本文提出了基于退化过程一致性的鉴别方法，包括基于退化过程的篡改网络和退化一致性鉴别网络。篡改网络模拟图像退化，生成退化特征不一致的伪造图像，为模型提供训练样本。而鉴别网络则从空间域和频域挖掘退化一致性线索，以提高模型的鉴别准确性。该方法即使在缺乏伪造样本的情况下，也能通过合成退化不一致的训练样本来有效构建鉴别模型。实验结果验证了所提方法的有效性。

       4. 面对新型伪造人脸数据难获取和人脸语义分析不足的问题，本文提出基于面部单元一致性的伪造人脸鉴别方法。面部形态学表明，真实人脸中不同面部单元区域之间表现出自然协调性。而在伪造人脸中，这种协调性通常会受到破坏。基于此发现，本文提出基于面部单元一致性的鉴别方法，包括面部单元关系学习网络和篡改面部单元预测模块。前者旨在学习面部动作单元之间的相互作用及其一致性模式。后者通过在图像层面和特征层面模拟伪造操作，产生面部单元不一致的训练样本，使模型对该线索更加敏感。该方法深入学习面部单元之间的复杂关系，即使仅用真实人脸图像进行训练，也能构建有效的鉴别模型，展现了良好的泛化能力。实验结果验证了该方法的有效性。

       综上所述，本文深入探讨了伪造图像鉴别领域所面临的关键挑战，并提出了若干创新性方法以提高伪造图像鉴别的准确性和泛化能力，为保障数字图像的真实性提供了新方法和新视角。

       Under the propulsion of digitalization and multimedia technology, images have become a crucial medium for information transmission and creative expression. The development of advanced image editing and image generation technologies has simplified the creation of lifelike visual works, advancing industries such as film and virtual reality. However, these technological advancements have also made image forgery easier and more convenient, blurring the line between authenticity and fabrication. Meticulously crafted forged images, especially those of fake faces, can easily mislead the public, causing panic and misunderstanding, and even be used for malicious purposes such as political manipulation and fraud, posing a serious threat to individual and societal security. Therefore, as technology becomes more widespread and the risk of misuse increases, developing effective forensic technologies to verify the authenticity of images has become a necessary and urgent task.

       Image forgery forensics faces multiple challenges, necessitating responses to different forms of forgery, including entirely computer-generated images and partially manipulated tampered images. Despite the effectiveness of existing forensic methods, limitations remain in the face of increasingly sophisticated forgery techniques: (1) In the identification of computer-generated images, existing datasets are of poor quality, and methods focus on global features, neglecting crucial detail differences; (2) In fake face detection, existing methods focus too much on specific forgery traces and overlook the common features of different forgery techniques; (3) Novel fake face data are difficult to acquire, making it challenging for models to undergo targeted training; (4) Existing methods fail to fully utilize the interconnectivity among facial units, affecting the in-depth understanding of facial characteristics. To address these issues, this paper proposes corresponding innovative methods, with the research content summarized as follows:

       1. Addressing the limitations of datasets and models in the identification of computer generated images, this paper constructs a Large-Scale Computer-Generated Image Benchmark (LSCGB) and proposes a computer-generated image identification method based on texture consistency. The LSCGB dataset is characterized by its large scale, high sample diversity, and low category bias, offering a more challenging testing platform for the field. Building on this, by deeply analyzing the texture feature differences between computer-generated and real images, a discrimination model based on texture consistency is proposed. This model enhances the texture information in the input features and integrates texture features at multiple scales to accurately distinguish between computer-generated and real images. Compared to existing methods and through ablation study results, the proposed method demonstrates effectiveness in terms of generalization ability across datasets and robustness to image post-processing operations.

       2. Facing the issue that existing fake face identification methods overlook common features, this paper proposes a fake face identification method based on noise consistency. The face forgery process introduces significant inconsistencies in noise between the face and the background. Based on this discovery, the method proposed in this paper includes two core modules: noise consistency enhancement and multi-scale noise consistency learning. The former enhances the feature representation capability by amplifying the noise differences between the face and background, while the latter calculates noise consistency between local areas at multiple scales and integrates information through a multi-scale feature integration strategy to improve the accuracy of forgery identification. After extensive testing on multiple datasets, the proposed method has achieved performance improvements in terms of accuracy and generalizability in fake face identification.

        3. To address the challenge of acquiring novel fake face data, this paper proposes a fake face identification method based on the consistency of the degradation process, reducing the reliance on fake face training data. Images undergo different degradation processes during capture and transmission, carrying varying degradation information. Forgery algorithms often splice faces from different sources with backgrounds, leading to inconsistencies in the degradation processes among these regions. Based on this observation, the paper introduces a discrimination method based on the consistency of the degradation process, including a tampering network based on the degradation process and a degradation consistency discrimination network. The tampering network simulates image degradation to generate fake images with inconsistent degradation features, providing training samples for the model. The discrimination network then mines degradation consistency clues from both the spatial and frequency domains to improve the model's discrimination accuracy. This method can effectively construct a discrimination model by synthesizing training samples with inconsistent degradation, even in the absence of fake samples. Experimental results validate the effectiveness of the proposed method.

       4. Addressing the difficulties in acquiring novel fake face data and the insufficiency of facial semantic analysis, this paper proposes a fake face identification method based on the consistency of facial units. Facial morphology demonstrates that different facial unit areas in real faces exhibit natural cohesiveness. In fake faces, this cohesiveness is often compromised. Based on this discovery, the paper introduces a discrimination method based on the consistency of facial units, including a facial unit relationship learning network and a tampered facial unit prediction module. The former aims to learn the interactions between facial action units and their consistency patterns. The latter generates training samples with inconsistent facial units by simulating forgery operations at both the image and feature levels, making the model more sensitive to this clue. This method delves into the complex relationships between facial units and can build an effective discrimination model even with training solely on real face images, demonstrating good generalization capabilities. Experimental results validate the effectiveness of this method.

        In summary, this paper thoroughly investigates the key challenges faced in the field of image forgery forensics and proposes several innovative methods to enhance the accuracy and generalizability of forgery image identification, offering new approaches and perspectives for ensuring the authenticity of digital images.