forensic evidence extraction

Existing image manipulation detection methods often suffer from poor robustness, fragmented forensic evidence, and limited generalization across manipulation types. To address these limitations, this work proposes an adaptive multi-path forensic framework that dynamically selects the optimal analysis pathway for each input image through a routing mechanism and integrates complementary evidence from both handcrafted and deep learning features. By moving beyond conventional single-model analysis and fixed fusion strategies, the proposed approach significantly improves detection and localization accuracy across diverse manipulation scenarios while maintaining a favorable balance between model performance and interpretability, thereby demonstrating enhanced robustness and cross-type generalization capability.

To address the limited zero-shot generalization capability in AI-generated image detection—where conventional methods rely on training data from known generators and thus fail to identify images produced by unseen models—this paper proposes the Forensic Self-Descriptor (FSD) framework. FSD employs self-supervised learning exclusively on authentic images, modeling multi-scale microstructures to extract generator-agnostic forensic residuals intrinsic to synthetic image formation, yielding compact, image-level representations. Crucially, it requires no synthetic images or generator priors. Consequently, FSD enables zero-shot detection, open-set source attribution, and unsupervised source clustering. Extensive evaluations across multiple benchmarks demonstrate substantial improvements over state-of-the-art methods, with strong cross-architecture generalization and robustness. To our knowledge, FSD is the first approach to achieve truly prior-free, universal AI image provenance tracing.

Current multimodal large language models (MLLMs) excel at semantic reasoning but struggle to perceive low-level forensic artifacts, leading to inaccurate tampering localization. To address this, we propose a two-stage forensics-driven framework: first, an enhanced LLaVA generates semantic-level tampering proposals; second, a forensic correction module injects multi-scale filtering–extracted low-level forgery features into SAM’s image embeddings to jointly optimize segmentation. This work establishes the first explicit alignment and fusion of semantic understanding with forensic features. Our method achieves state-of-the-art detection and localization accuracy across multiple benchmarks, significantly improving robustness against complex manipulations—including splicing and copy-move—and enhancing cross-domain generalization capability.

Existing video forensic methods typically target a single manipulation type (e.g., deepfakes or inpainting), rendering them inadequate for real-world scenarios where manipulation types are unknown and often co-occur. This paper introduces the first end-to-end, multi-purpose video forensic network capable of jointly detecting diverse manipulations—including deepfakes, inpainting, splicing, and editing—without prior knowledge of the manipulation type. Our method features a novel multi-scale hierarchical Transformer module that jointly models spatiotemporal anomalies and precisely localizes forged regions of arbitrary shape and size across scales. Additionally, it integrates multimodal forensic cues with multi-scale spatiotemporal features. Evaluated on a comprehensive multi-manipulation benchmark, our approach achieves state-of-the-art performance, while also matching or surpassing specialized detectors on single-type manipulation tasks—demonstrating significantly improved generalization and practical applicability.

This work proposes the first extended reality (XR)-based animation authoring framework tailored for criminal investigation applications, addressing the challenge that existing animation tools are ill-suited for non-expert forensic personnel to efficiently reconstruct and validate dynamic crime scenes. Developed through close collaboration with criminology experts, the system features an intuitive, low-barrier interface that enables users without 3D modeling experience to rapidly create and observe animated reconstructions of criminal events. Evaluated by 18 participants—including six trained criminology professionals—the system demonstrates high task completion rates and strong usability scores in character animation tasks, effectively supporting diverse use cases such as hypothesis validation, case presentation, situational understanding, and training.