Transparency Attacks: How Imperceptible Image Layers Can Fool AI Perception · HackerPapers discussionshttps://hackerpaper.kaipfister.de/p/2401.15817/Collected HN comments and submissions linking this paper.enMon, 02 Nov 2026 11:16:22 GMTTransparency Attacks: How Imperceptible Image Layers Can Fool AI Perception — mentioned by squarefoothttps://news.ycombinator.com/item?id=40561220urn:hackerpapers:mention:2401.15817:40561220Mon, 03 Jun 2024 10:29:04 GMT<p>Because you know, AI never gets an image wrong, right? https://arxiv.org/pdf/2401.15817 Quite long paper, but rich in examples that speak for themselves, like AI mistaking a Starbucks logo for a Ferrari one, or an airplane for a baby stroller, etc. CSAM producers will always find a way to poison images or use other tricks to conceal them. There was one in fact years ago that involved appending a .rar archive to a .jpg image without renaming the image (including the .jpg suffix): an image viewer would correctly show the image ignoring the following data, while a .rar unpacker would ignore…</p><p><a href="https://hackerpaper.kaipfister.de/p/2401.15817/">Paper and discussion history</a>. A mention is not an endorsement.</p>cs.CV