AI Content Watermarking Techniques

📝 Announcing our #CVPR 2026 paper that introduces the first watermarking technique that is safe against visual paraphrase attacks, while remaining distortion-free (preserving visual fidelity) and model-agnostic across AI image generators. 🔹 𝐕𝐢𝐬𝐮𝐚𝐥 𝐏𝐚𝐫𝐚𝐩𝐡𝐫𝐚𝐬𝐞 𝐀𝐭𝐭𝐚𝐜𝐤 𝐒𝐚𝐟𝐞 𝐚𝐧𝐝 𝐃𝐢𝐬𝐭𝐨𝐫𝐭𝐢𝐨𝐧-𝐅𝐫𝐞𝐞 𝐈𝐦𝐚𝐠𝐞 𝐖𝐚𝐭𝐞𝐫𝐦𝐚𝐫𝐤𝐢𝐧𝐠 𝐓𝐞𝐜𝐡𝐧𝐢𝐪𝐮𝐞 𝐟𝐨𝐫 𝐀𝐈-𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐞𝐝 𝐈𝐦𝐚𝐠𝐞𝐬 🔹 In collaboration with Vishwakarma Institute of Information Technology, Indraprastha Institute of Information Technology, Delhi, BITS Pilani, Hyderabad Campus, AI at Meta, and University of South Carolina 🔹 Paper: https://lnkd.in/gazGcD_k ✍🏼 Authors: Shreyas Dixit, Ashhar Aziz, Shashwat Bajpai, Vasu Sharma, Aman Chadha, Vinija Jain, Dr. Amitava Das ➡️ 𝐊𝐞𝐲 𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬 𝐨𝐟 𝐏𝐄𝐂𝐂𝐀𝐕𝐈: 🧩 𝐍𝐨𝐧-𝐌𝐞𝐥𝐭𝐢𝐧𝐠 𝐏𝐨𝐢𝐧𝐭𝐬 (𝐍𝐌𝐏𝐬) 𝐟𝐨𝐫 𝐑𝐨𝐛𝐮𝐬𝐭 𝐄𝐦𝐛𝐞𝐝𝐝𝐢𝐧𝐠: Identifies semantically stable regions that survive visual paraphrasing and embeds watermarks precisely where they are least likely to be altered. 🔐 𝐌𝐮𝐥𝐭𝐢-𝐂𝐡𝐚𝐧𝐧𝐞𝐥 𝐅𝐫𝐞𝐪𝐮𝐞𝐧𝐜𝐲-𝐃𝐨𝐦𝐚𝐢𝐧 𝐖𝐚𝐭𝐞𝐫𝐦𝐚𝐫𝐤𝐢𝐧𝐠: Distributes watermark signals across channels and patches in Fourier space, significantly improving resilience to compression, noise, regeneration, and paraphrase attacks. 🛡️ 𝐍𝐨𝐢𝐬𝐲 𝐁𝐮𝐫𝐧𝐢𝐬𝐡𝐢𝐧𝐠 & 𝐑𝐚𝐧𝐝𝐨𝐦 𝐏𝐚𝐭𝐜𝐡𝐢𝐧𝐠: Defends against reverse-engineering of salient regions and targeted removal by adversaries. ⚖️ 𝐀𝐝𝐚𝐩𝐭𝐢𝐯𝐞 𝐄𝐧𝐡𝐚𝐧𝐜𝐞𝐦𝐞𝐧𝐭 𝐟𝐨𝐫 𝐋𝐨𝐰 𝐃𝐢𝐬𝐭𝐨𝐫𝐭𝐢𝐨𝐧: Blends watermarked and original images to maintain high PSNR and SSIM while preserving strong detectability. 🌍 𝐌𝐨𝐝𝐞𝐥-𝐀𝐠𝐧𝐨𝐬𝐭𝐢𝐜 𝐚𝐧𝐝 𝐄𝐱𝐭𝐞𝐧𝐬𝐢𝐯𝐞𝐥𝐲 𝐄𝐯𝐚𝐥𝐮𝐚𝐭𝐞𝐝: Demonstrated across Stable Diffusion variants, DALL-E 3, and Midjourney, outperforming prior watermarking methods under visual paraphrase attacks. With AI-generated content rapidly becoming the majority of online media, we hope PECCAVI contributes to a more trustworthy and accountable generative AI ecosystem. #artificialintelligence #genai #responsibleai #research

Can you tell the difference between human-written language and AI-generated text?🤔 To solve this problem we need watermarks!📃 Researchers at the University of Maryland(https://lnkd.in/gG7gpZpJ) created a way for us to modify LLMs such that a watermark would automatically be applied to any content that LLM generates. This allows us to run a test for this watermark to identify synthetic content in the wild. A watermark is a hidden pattern in text that is imperceptible to humans, but when the text is statistically analyzed it allows us to identify synthetic content. The watermark they created can be identified in as little as 25 tokens and has negligible impact on text quality. The watermark works by selecting a randomized set of secret “green” tokens before a word is generated, and then softly incentivizing the LLM to use those green tokens by slightly nudging the output word probabilities during sampling. The more "green" tokens found in a chunk of text the higher the probability it was generated by an LLM. The challenge here is that in order to apply this watermark the company owning the LLM (OpenAI, Cohere, Anthropic etc.) needs to promote the use of these random secret "green" tokens by slightly increasing their probability of being generated. Yet, another problem is that the higher this "green" token probability the easier the watermark will be to detect however, this also lowers the quality of the text overall.

#TuesdayPaperThoughts Edition 14: Watermarking GenAI This week's #TuesdayPaperThoughts highlights "Scalable watermarking for identifying large language model outputs", latest paper from Google DeepMind on watermarking generative text at scale without affecting generation quality and latency. Key insights: 1️⃣ SynthID only modifies the sampling procedure. It adjusts the probability score of each predicted token where there's a range of different tokens to choose from without compromising the quality, accuracy and creativity of the output. 2️⃣ This process is repeated throughout the generated text. The chosen words along with the adjusted probability scores are considered the watermark. The watermarking is classified through a scoring and thresholding function at test time. 3️⃣ SynthID has neutral results on the human side-by-side ratings indicating no change in LLM capabilities. Around 20M Gemini response feedbacks were also assessed to confirm. As far as I know, this is the first deployment of GenAI watermarking in production and will be very helpful for the information ecosystem. Research Credits: Sumanth Dathathri, Abigail See, Sumedh Ghaisas, Po-Sen Huang, Rob McAdam, Johannes Welbl, Vandana Bachani, Alex Kaskasoli, Robert Stanforth, Tatiana Matejovicova, Jamie Hayes, Nidhi Vyas, Majd Al Merey, Jonah Brown-Cohen, Rudy Bunel, Borja Balle, Ali Taylan Cemgil, Zahra A., Kitty Stacpoole, Ilia Shumailov, Ciprian Baetu, Sven Gowal, Demis Hassabis, Pushmeet Kohli Paper Link: https://lnkd.in/gprnNyhU

Big AI news: Google now officially watermarks all AI text outputs, and today published a paper in Nature Magazine detailing how it works! Watermarking of AI-generated content is critically important, because without it, it is very difficult to tell what is authentic and what is fake. This paper is big news because there is a pitched battle taking place about text watermarking right now behind the scenes in places like Brussels, Sacramento and Washington. Although California and the EU now have laws requiring watermarking of AI-generated content, they haven't yet come into force, and some AI companies are fighting to exempt text, and have the laws only apply to images, audio and video. The companies fighting against text watermarking include OpenAI, Microsoft and their friends at industry lobbying associations NetChoice and TechNet. A Wall Street Journal expose (linked in comments) in August uncovered evidence that OpenAI actually has developed this technology in house, tested it on users, found that it was effective, and chose not to launch it after research showed that 30% of their paying users would use ChatGPT less if it had watermarks and competitors didn't. Hopefully this big step forward from Google DeepMind will be a push towards getting OpenAI, Anthropic, Meta and others to follow suit. If not, I'll be working hard to get laws passed that make them do it. Abstract: Large language models (LLMs) have enabled the generation of high-quality synthetic text, often indistinguishable from human-written content, at a scale that can markedly affect the nature of the information ecosystem1,2,3. Watermarking can help identify synthetic text and limit accidental or deliberate misuse4, but has not been adopted in production systems owing to stringent quality, detectability and computational efficiency requirements. Here we describe SynthID-Text, a production-ready text watermarking scheme that preserves text quality and enables high detection accuracy, with minimal latency overhead. SynthID-Text does not affect LLM training and modifies only the sampling procedure; watermark detection is computationally efficient, without using the underlying LLM. To enable watermarking at scale, we develop an algorithm integrating watermarking with speculative sampling, an efficiency technique frequently used in production systems5. Evaluations across multiple LLMs empirically show that SynthID-Text provides improved detectability over comparable methods.... Kudos to the team of authors at Google Deepmind: Sumanth Dathathri, Abigail See, Sumedh Ghaisas, Po-Sen Huang, Rob McAdam, Johannes Welbl, Vandana Bachani, Alex Kaskasoli, @Robert Stanforth, Tatiana Matejovicova, Jamie Hayes, Nidhi Vyas, Majd Al Merey, Jonah Brown-Cohen, Rudy Bunel, Borja Balle, Ali Taylan Cemgil, Zahra Ahmed, Kitty Stacpoole, Ilia Shumailov, Ciprian Baetu, Sven Gowal, Demis Hassabis (recent winner of The Nobel Prize!) & Pushmeet Kohli.