🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Mirror, Mirror on the wall, who is the biggest Panic-creator of them all? Inspired by Nirit Weiss-Blatt, PhD, check out our Taxonomy of AI Panic Facilitators and those fighting against the fearmongering. Who have we forgotten to add? Let us know! ⬇️

I’m quite used to the cruelty students can face when they apply for a US visa but this one broke me. We offered admission to a stellar, talented & hardworking student. After months of work and hundreds of dollars, an embassy officer saw him for 5 mins & said no. why? …

Earlier today, The Minister for Electronics & IT, Ashwini Vaishnaw introduced the long-awaited, much-iterated, & disappointing Digital Personal Data Protection Bill, 2023 in the Lok Sabha. #SaveOurPrivacy Check out our first read of the Bill ⬇️ 1/10 internetfreedom.in/iffs-first-rea…

Exclusive: the government has conceded that technology to scan encrypted messages does not currently exist and powers in the online safety bill could not be used until “technically feasible” w/Anna Gross on.ft.com/486YC67

SO good -- the best follow the money reporting on who's behind the global attack on digital privacy yet. TLDR: it's law enforcement x AI companies posing as NGOs w a commercial interest in selling scammy mass scanning tech. Deeply cynical, deeply shady. balkaninsight.com/2023/09/25/who…

The Commission is using microtargetted advertising to push their CSAM proposal. Besides being antidemocratic (and apparently illegal due to the target categories) this is yet another sign that they have no real argument why the regulation should pass dannymekic.com/202310/undermi…

🤔 Did an LLM read your book 📘? In a new arXiv.org preprint, we propose the document-level membership inference task for LLMs. Using only the LLM output, we can accurately tell if a document was seen during training (AUC=0.856) arxiv.org/abs/2310.15007. A short 🧵

🔥 Great piece in Le Monde on our method to detect whether an LLM has seen a book 📚 during training lemonde.fr/sciences/artic… Paper: arxiv.org/abs/2310.15007 Work with Matthieu Meeus, Shubham Jain, and Marek Rei at Imperial College London

pluralistic.net/2023/12/08/pla… "If buying isn't owning, piracy isn't stealing" by Cory Doctorow NONCONSENSUAL BLUE TICK is a great essay shedding light on why we need regulations like right to repair. Last month, I had to sell my old tablet because there was no way to update/change its OS and repurpose it.

Ross Anderson Professor Ross Anderson, FRS, FREng Dear friend and treasured long term campaigner for privacy and security, Professor of Security Engineering at Cambridge University and Edinburgh University, Lovelace Medal winner, has died suddenly at home in Cambridge.

I was privileged to have Ross as my PhD Thesis Examiner in October last year. He was a pioneer in the field of privacy and security, but still humane. He made me feel equal while we discussed my thesis and his thoughts on it. Would always be grateful for his contributions.

Statement, signed by 250+ researchers, warning that the modifications of the Regulation to detect CSAM proposed by the EU presidency dont solve the issues pointed out by experts. It still introduces societal risks without solving the CSA problem. csa-scientist-open-letter.org

I’m in Philadelphia for USENIX Security! Hit me up if you’d like to discuss privacy/copyright of LLMs and come check out our talk on document-level membership inference for LLMs on Thursday morning. In the meantime I had to try my first cheesesteak 🧀🥩

Long fascinating thread by Matthieu Meeus on document-level membership inference on LLMs: - early hopes 🥳 (2 concurrent papers and more methods since) - the serious distribution shift concerns 🤯 - and potential ways forward for the field incl. controlled insertion of traps 🪤

This talk from Nicholas Carlini goes *hard* - the story of what went wrong with adversarial robustness, and warnings for what pitfalls to avoid in alignment research. It's an excellent talk, and the recording is now out! Highly recommended

Upon popular demand, we release our source code for creating a deduplicated version of the ImageNet dataset, used in our USENIX Security ‘22 paper evaluating the robustness of perceptual hashing-based client-side scanning to detection avoidance attacks. github.com/computationalp…

Are membership inference attacks (MIAs) against LLMs rushing nowhere? 🏃‍ ➡️ In a new SoK, we look at how things have evolved recently, show popular evaluation setups to be flawed, and examine solutions going forward.