Abstract:

With the development of multimedia editing software and generative neural networks, the reliability of digital images is being continuously eroded. As an emerging forensic technique for provenance analysis, image attribution retraces the trustworthy source of the image under analysis and visualizes the editorial changes in such image. Thus, it can effectively combat malicious manipulation, assisting users to form correct judgments on image information. However, current image attribution methods are not sufficiently robust to the geometric transformations or signal corruptions in modern cyberspace, especially for images that contain multiple distortions. For this gap, an image attribution method with multi-distortion robustness was proposed. The method was based on an orthogonal and covariant image local representation strategy with robustness to multiple geometric transformations or signal corruptions. Two fast implementations were designed for sparse and dense representation tasks, respectively. The resulting image attribution method was able to efficiently retrace near-duplicate source in a trusted database, correct the geometric pose, and visualize potential tampering regions. In such process, the proposed method was robust to various benign transformations while maintaining sensitivity to subtle content manipulation. Simulation results show that the proposed image attribution method exhibits better forgery detection robustness and overall accuracy, as well as better feature compactness and implementation cost.

Key words: geometric invariance, image attribution, perceptual hashing, forgery detection, near-duplicate retrieval