Shen, Qi, Gong, Yujia, Zhao, Yue, Liu, Tong · Nov 19, 2026
A Wolf in Sheep's Pixels: Visual Camouflage Attacks via HDR Gain Maps
Markdown · JSON · Original source
Qi Shen, Yujia Gong, Yue Zhao and Tong Liu introduce DRJanus, a visual camouflage attack that uses the difference between standard dynamic range (SDR) processing and high dynamic range (HDR) display. Their abstract says automated systems often read an image’s SDR base, while HDR viewers reconstruct its appearance with an auxiliary gain map. DRJanus inverts that reconstruction to place two independently chosen color images in one JPEG, without querying or optimizing against the target model. The authors report that the SDR base remains unchanged and that HDR targets reach an average PSNR of 43.03 dB across 400 images. On a of 100 sexual-content images, they report evasion of all six tested commercial moderation services and nine vision-language models. Across five agents, they report attack success rates of 31.67% to 100% for moderation evasion and 95% to 100% for task manipulation. They also explore cross-view consistency checks as a mitigation. The paper’s central warning is that securing what a machine reads does not secure what a user sees. The abstract does not establish how well the proposed checks work beyond saying they were explored.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 19, 2026 (2026-11-19).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-20T08:46:16.294Z. First observed: 2026-11-20T05:07:57.330Z. Retrieved: 2026-11-20T08:33:08.161Z.
Referenced by
Recorded citations in public readings. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The supplied evidence is the abstract; it does not provide details of the experiments or report how effective the cross-view consistency checks are. The authors report results for the stated and tested services, models and agents; the evidence does not establish performance beyond those settings.
Open question
How effective are the cross-view consistency checks the authors explored at mitigating DRJanus?
Source-stated
The authors introduce DRJanus, which exploits a difference between SDR processing and HDR display.
We introduce DRJanus, a visual camouflage attack that exploits the gap between standard dynamic range (SDR) processing and high dynamic range (HDR) display.
Source-stated
The authors say DRJanus embeds two independently chosen color images in one JPEG, without querying or optimizing against the target model, and preserves the SDR base exactly.
inverts this reconstruction to embed two independently chosen color images in a single JPEG, without querying or optimizing against the target model. It preserves the chosen SDR base exactly
Source-stated
The authors report evasion of all six tested commercial moderation services and nine vision-language models on a benchmark of sexual-content images.
On a benchmark of 100 sexual-content images, DRJanus evades detection by all six tested commercial moderation services and nine vision-language models.
Source-stated
The authors report attack success rates across five agents and say they explore cross-view consistency checks as a mitigation.
Across five agents, each evaluated on 120 task cases per threat setting, attack success ranges from 31.67% to 100% for moderation evasion and from 95% to 100% for task manipulation. We also explore cross-view consistency checks as a mitigation.
Source and contributor context
Shen, Qi, Gong, Yujia, Zhao, Yue, Liu, Tong · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: da915cfb-f71a-43f1-867b-beacb7b28cdc.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-20T08:46:16.294Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 779483afebda85566cf073718afc68007f6ac83d64726309f484343aedecbfa5. Source text hash: a0bfa54bc5c38ca3c4767a7a0325653cdac6786cb854304e81b761e883d9fc43.
Revision and source-check history
Version 1 · 2026-11-20T08:46:16.294Z
A Wolf in Sheep's Pixels: Visual Camouflage Attacks via HDR Gain Maps
Qi Shen, Yujia Gong, Yue Zhao and Tong Liu introduce DRJanus, a visual camouflage attack that uses the difference between standard dynamic range (SDR) processing and high dynamic range (HDR) display. Their abstract says automated systems often read an image’s SDR base, while HDR viewers reconstruct its appearance with an auxiliary gain map. DRJanus inverts that reconstruction to place two independently chosen color images in one JPEG, without querying or optimizing against the target model. The authors report that the SDR base remains unchanged and that HDR targets reach an average PSNR of 43.03 dB across 400 images. On a of 100 sexual-content images, they report evasion of all six tested commercial moderation services and nine vision-language models. Across five agents, they report attack success rates of 31.67% to 100% for moderation evasion and 95% to 100% for task manipulation. They also explore cross-view consistency checks as a mitigation. The paper’s central warning is that securing what a machine reads does not secure what a user sees. The abstract does not establish how well the proposed checks work beyond saying they were explored.
Original source
cs.AI · Article
Source details
Original source date: Nov 19, 2026 (2026-11-19).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-20T08:46:16.294Z. First observed: 2026-11-20T05:07:57.330Z. Retrieved: 2026-11-20T08:33:08.161Z.
Limits
The supplied evidence is the abstract; it does not provide details of the experiments or report how effective the cross-view consistency checks are. The authors report results for the stated and tested services, models and agents; the evidence does not establish performance beyond those settings.
Open question
How effective are the cross-view consistency checks the authors explored at mitigating DRJanus?
Source-stated
The authors introduce DRJanus, which exploits a difference between SDR processing and HDR display.
We introduce DRJanus, a visual camouflage attack that exploits the gap between standard dynamic range (SDR) processing and high dynamic range (HDR) display.
Source-stated
The authors say DRJanus embeds two independently chosen color images in one JPEG, without querying or optimizing against the target model, and preserves the SDR base exactly.
inverts this reconstruction to embed two independently chosen color images in a single JPEG, without querying or optimizing against the target model. It preserves the chosen SDR base exactly
Source-stated
The authors report evasion of all six tested commercial moderation services and nine vision-language models on a benchmark of sexual-content images.
On a benchmark of 100 sexual-content images, DRJanus evades detection by all six tested commercial moderation services and nine vision-language models.
Source-stated
The authors report attack success rates across five agents and say they explore cross-view consistency checks as a mitigation.
Across five agents, each evaluated on 120 task cases per threat setting, attack success ranges from 31.67% to 100% for moderation evasion and from 95% to 100% for task manipulation. We also explore cross-view consistency checks as a mitigation.
Source and contributor context
Shen, Qi, Gong, Yujia, Zhao, Yue, Liu, Tong · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: da915cfb-f71a-43f1-867b-beacb7b28cdc.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-20T08:46:16.294Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 779483afebda85566cf073718afc68007f6ac83d64726309f484343aedecbfa5. Source text hash: a0bfa54bc5c38ca3c4767a7a0325653cdac6786cb854304e81b761e883d9fc43.
passed · 2026-11-20T08:46:16.294Z · anthropic/claude-opus-5.5
Publication: da915cfb-f71a-43f1-867b-beacb7b28cdc
attribution — pass: The reading names the four listed authors and attributes the method, results and warning to them or their abstract. It uses phrasing such as 'they report' and 'the authors report', so the results are not presented as independently established.
Authors:Qi Shen, Yujia Gong, Yue Zhao, Tong Liu
summary — pass: The reading leads with the actors and the DRJanus attack. It explains the SDR/HDR gap in plain terms: machines read the base image, while HDR viewers rebuild what the person sees. It keeps the paper's central warning in the authors' words. Terms such as PSNR, vision-language models and agents are left for tooltips, but this does not block understanding of the main idea. The question is a fair unresolved issue, since the abstract only says mitigation was explored.
securing what a machine reads does not secure what a user sees
date — pass: The originalPublished date of 2026-11-19 matches the arXiv v1 submission date. The feed date of 2026-11-20 is later and was correctly not used.
[Submitted on 19 Nov 2026]
support — pass: Every figure is source-stated and correctly scoped: 43.03 dB average PSNR over 400 images, 100 sexual-content images, six moderation services, nine vision-language models, five agents, 31.67%–100% and 95%–100%. 'SDR base remains unchanged' matches 'preserves... exactly'. The limitations accurately note that mitigation effectiveness is not reported. No benefits or costs are invented.
It preserves the chosen SDR base exactly and reconstructs HDR targets at an average PSNR of 43.03 dB across 400 images.
media — pass: The selected media is the verified arXiv abstract page for this paper, which is the direct original source. No better source media was found.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference