VeriTrace: A Multitask Disaster Dataset for Real–Fake Verification and Source Tracing

Abstract

The rapid increase in AI-generated images is affecting not only face-focused domains like face swapping and media manipulation but also non-facial content such as scenes and objects. Yet, one crucial area that remains underexplored is disaster deepfakes. Natural disasters such as earthquakes, floods, and wildfires occur unpredictably, and the widespread use of social media has amplified the circulation of fake disaster images that can mislead the public about the severity and consequences of these events. Detecting whether a disaster image is real or AI-generated is challenging due to the semantic complexity of disaster scenes, the scarcity of labeled data, and the difficulty of identifying images produced by previously unseen generative models. We argue that the limited diversity of existing training data is a key barrier to building robust detection systems. To address this, we introduce VeriTrace, a large-scale, in-the-wild multitask dataset curated via a CLIP-based pipeline containing AI-generated disaster images from 3,191 generative models. VeriTrace supports both deepfake detection and source tracing. Our benchmarks confirm that disaster-domain deepfakes are significantly harder to detect than human-domain. Furthermore, we identify an identification-attribution gap: while high-fidelity commercial images frequently evade standard detectors, they possess distinct architectural “fingerprints” that make them highly susceptible to source tracing. VeriTrace and its associated code will be released upon acceptance.