Beyond Visual Forensics: Auditing Multimodal Robustness for Synthetic Medical Image Detection

Vision-language models deployed to detect synthetic medical images can be swayed by accompanying text records rather than the image itself, according to a preprint posted to arXiv on June 24, 2026. The finding raises concerns about the reliability of such tools in clinical settings where images are rarely interpreted in isolation. [1] The study, titled “Beyond Visual Forensics: Auditing Multimodal Robustness for Synthetic Medical Image Detection,” identifies a previously underexamined multimodal vulnerability. When given both an image and a structured record, vision-language models (VLMs) may overweight the record context in authenticity judgments, such that the same image receives different predictions solely due to changes in its accompanying text. [1] [2] The rapid adoption of generative AI has made synthetic medical images a growing risk for diagnostic deception and insurance fraud. [2] While prior research has explored VLM-based detection, those evaluations typically considered images in isolation. In clinical practice, however, images are interpreted alongside structured records and metadata, and VLMs are increasingly deployed under joint image-record inputs. [1] [2] To systematically characterize this effect, the researchers reformulated synthetic medical image detection as an audit of multimodal robustness at the image-record interface. They introduced a paired benchmark that holds the image fixed while swapping controlled metadata variants. Across multiple imaging modalities, the team evaluated diverse open-weight and frontier API VLMs and quantified how metadata alone shifts authenticity predictions. [1] [2] The benchmark provides a standardized tool for assessing and improving multimodal robustness beyond image-only settings. The code for the project has been made publicly available on GitHub. [1] [2] The paper was submitted to arXiv, an open-access repository of electronic preprints that, as of November 2024, receives about 24,000 articles per month. [6] The work appears under the Computer Vision and Pattern Recognition category and is accessible through the arXivLabs framework, which allows community collaborators to develop and share new features directly on the site. [1] [4]