HDR 'Gain Map' JPEGs Can Hide Content From Automated Moderation, Researchers Say
Researchers say they have found a way for a single HDR JPEG image to slip past automated image safety checks while showing different content to a person on an HDR-capable device, according to a new arXiv preprint and a public demo released this week.
The paper, titled “A Wolf in Sheep’s Pixels: Visual Camouflage Attacks via HDR Gain Maps,” describes a technique the authors call DRJanus. Its core claim is simple: “One JPEG can show different content to an automated system and a person,” as the project’s demo page puts it. In the paper, the researchers write that the work exposes “a gap in visual trust: securing what a machine reads does not secure what a user sees.”
In plain terms, the split happens because some image-processing systems may inspect only a picture’s standard dynamic range, or SDR, base layer. But an HDR-capable viewer can reconstruct a different final image using extra metadata known as a gain map, which is used to expand brightness and contrast for high dynamic range displays. According to the paper, that can create “one JPEG, two views”: a benign-looking SDR version for moderation systems and a different HDR rendering for a human viewer.
The authors report several strong results, all from their own experiments. In a benchmark using 100 sexual-content target images, they say DRJanus images evaded detection by all six commercial moderation services they tested: OpenAI Moderation, Amazon Rekognition, Azure AI Content Safety, Google Cloud Vision SafeSearch, Tencent Cloud Image Moderation and Alibaba Cloud Image Moderation. The paper also says the method bypassed nine vision-language models that detected the corresponding HDR control images but not the DRJanus JPEGs. In tests on five multimodal agent systems, the authors report moderation-evasion success rates ranging from 31.67% to 100%, and task-manipulation success rates from 95% to 100%, across 120 cases in each category per system. They also report an average PSNR, a common image fidelity metric, of 43.03 dB across 400 RGB reconstructions, which they present as evidence that the HDR target image can be recovered with high quality.
The preprint was posted to arXiv on Nov. 19 as arXiv:2611.25469v1. Its authors are Qi Shen, Yujia Gong, Yue Zhao and Tong Liu, whose affiliations are listed as the University of the Chinese Academy of Sciences and the Institute of Information Engineering. As with any arXiv paper, the findings should be treated as preliminary until they are independently reviewed and replicated. The authors say they responsibly disclosed the issue to affected vendors, and they write that one major social media platform confirmed the problem, rated it high severity and paid a bug bounty. But no independent public vendor advisories or press statements confirming that claim were found as of Nov. 20.
The idea is technically plausible because gain-map HDR is no longer a niche format. The paper points to ISO 21496-1:2025, a standard for gain-map metadata used in dynamic-range conversion. Similar approaches are already visible in mainstream consumer ecosystems, including Android Ultra HDR and Apple Adaptive HDR. That matters because it suggests the attack is aimed at real image pipelines used by phones, apps and online platforms, rather than a narrowly tailored lab setup. The paper also frames DRJanus as part of a broader class of split-view attacks, where a file presents one version to automated systems and another to human viewers.
The researchers also tested a possible defense: cross-view consistency checks that compare the SDR base image with the HDR reconstruction. They report that this detected all 600 tested attacks in their experiments. The paper also highlights a key limitation. The attack depends on platforms and viewers preserving and rendering the HDR gain map. If a service strips that metadata or converts the upload to plain SDR, the dual-view effect disappears.