Abstract: With their widespread use on social media, digital images have become core carriers of information dissemination. The rise of powerful and easy-to-use image editing software and generative artificial intelligence technologies has lowered the threshold for creation, while also providing a more covert method for malicious image tampering, leading to the accelerated spread of false information. Tampering leaves specific tampering features in the image that constitute the core basis of image tampering detection technology. In the face of increasingly complex and diverse tampering methods, most existing reviews focus on a single technical route and lack a systematic comparison and integrated analysis across different image tampering detection techniques. To this end, this study constructs a three-dimensional classification system of ″feature traceability—extraction method—detection task″, and categorizes image tamper detection technologies into two main classes based on tampering features: manual feature-based and deep learning feature-based approaches. Specifically, the following work is carried out: First, the classification framework is systematically reconstructed for manual feature-based image tampering detection techniques, consolidating the scattered manual features in traditional research into three major categories-camera system, pixel-level, and format-related features. An in-depth analysis of the physical mechanisms and improvement effects of the performance optimization strategies is conducted for 14 typical image tampering detection techniques, thereby compensating for the deficiency of existing reviews in the systematic analysis of manual feature-based methods. Second, deep learning feature-based image tampering detection techniques are architecturally organized, with a focused analysis on generative image tampering detection. Third, existing tampered image datasets are summarized and commented on with respect to their composition, characteristics, and limitations, thereby providing a basis for dataset selection. Finally, a summary and outlook on future research directions and development trends in this field are presented, and several key scientific issues that urgently need to be addressed are highlighted, thereby providing a useful reference for subsequent research.

Key words: image tampering detection, handcrafted feature, performance optimization strategy, deep learning feature, dataset with tampering image

摘要： 随着数字图像在社交媒体中的广泛应用，其已成为信息传播的核心载体。功能强大且易于使用的图像编辑软件与生成式人工智能技术的兴起，在降低创作门槛的同时，也为图像恶意篡改提供了更隐蔽的途径，导致虚假信息加速扩散。篡改行为会在图像中留下具有特异性的篡改特征，这构成了图像篡改检测技术的核心依据。面对日益复杂且多样化的篡改手段，现有综述多聚焦于单一技术路线，缺乏对图像篡改检测技术的系统性对比与整合分析。为此，本文构建“特征溯源—提取方式—检测任务”三维分类体系，基于篡改特征将图像篡改检测技术归纳为基于手工特征和基于深度学习特征两大类，并进行以下工作：首先，系统重构基于手工特征的图像篡改检测技术分类框架，将传统研究中分散的手工特征整合为相机系统特征、像素级特征与格式相关特征三大类，深入剖析14种典型图像篡改检测技术的性能优化策略、物理机理与改进效果，弥补现有综述在基于手工特征的图像篡改检测技术系统性分析方面的不足；然后，架构化梳理基于深度学习特征的图像篡改检测技术，并针对生成式图像篡改检测技术进行着重分析；接着，对现有篡改图像数据集的构成、特点及其局限性进行归纳与评述，为数据集选择提供可选择的依据；最后，总结展望该领域未来的研究方向和发展趋势，指出若干亟待解决的关键科学问题，以期为后续研究提供参考与借鉴。

关键词: 图像篡改检测, 手工特征, 性能优化策略, 深度学习特征, 篡改图像数据集

CLC Number: 

• TP309.2