A Gaussian mechanism with ε = 6 can be less private than one with ε = 8. This points to a problem with how we report privacy guarantees in machine learning. A thread 🧵

Interesting "Truly private AI" that seems like the first to deliver that, up to what current technology enables.

(Yes, you have to trust the opaque magic of Trusted Execution Environments, but it is hard to see how to realistically avoid that for now.)

confer.to

📣 Call for posters for @elsa-ai.eu TrustworthyAI4Health Workshop co-located w/ @embl.org AI & Biology Conference, on topics that advance reliable, clinically aligned AI systems across diverse data modalities and healthcare environments.

📅 Mar 9
📍 Heidelberg 🇩🇪
🔗 https://bit.ly/4pEK3OK

#EESAIBio

ELLIS Institute Finland
@ellisinstitute.fi
has an open call for postdocs (DL 9 Feb) www.ellisinstitute.fi/postdoc-recr...

There are 45 PIs with different topics to choose from, including privacy in machine learning with me!

Text: We're hiring! Principal Investigators in artificial intelligence and machine learning research. Below: ELLIS Institute Finland logo

1 month until deadline! Join us to build your own lab in AI + machine learning research. World-class resources incl. @lumi-supercomputer.eu, generous starting package & professorship affiliation with a university in the world’s happiest country! ➡️ www.ellisinstitute.fi/PI-recruit-2...
#hiring

How to ensure anonymity of AI systems? | Faculty of Science | University of Helsinki When training artificial intelligence systems, developers need to use privacy-enhancing technologies to ensure that the subjects of the training data are not exposed, new study suggests.

ELSA Board member @ahonkela.bsky.social contributed to the paper"Impact of Dataset Properties on Membership Inference Vulnerability of Deep Transfer Learning",presented at @neuripsconf.bsky.social'25

Article ➡️ www.helsinki.fi/en/faculty-s...

#MachineLearning #DifferentialPrivacy #PrivacyGuarantees

Extract of ICLR LLM use policy with the following text highlighted: "However, new this year, if LLMs played a significant role in research ideation and/or writing to the extent that they could be regarded as a contributor, then authors should describe the precise role of the LLM in the main body of the paper in a separate section on LLM usage."

@iclr-conf.bsky.social regarding your LLM policy, can you please confirm that one does not need to have an LLM usage section in the main body of the paper if LLMs did not play a significant role as defined here?

Happy to share the recap of the #health #privacy #session at #CAMDA2025 by our amazing team!

elsa-ai.eu/the-health-p...

#ELSAAI #research #challenge #safeandsecure #AI

$( varepsilon, δ)$ Considered Harmful: Best Practices for Reporting Differential Privacy Guarantees Juan Felipe Gomez, Bogdan Kulynych, Georgios Kaissis, Jamie Hayes, Borja Balle, Antti Honkela http://arxiv.org/abs/2503.10945 Current practices for reporting the level of differential privacy (DP) guarantees for machine learning (ML) algorithms provide an incomplete and potentially misleading picture of the guarantees and make it difficult to compare privacy levels across different settings. We argue for using Gaussian differential privacy (GDP) as the primary means of communicating DP guarantees in ML, with the full privacy profile as a secondary option in case GDP is too inaccurate. Unlike other widely used alternatives, GDP has only one parameter, which ensures easy comparability of guarantees, and it can accurately capture the full privacy profile of many important ML applications. To support our claims, we investigate the privacy profiles of state-of-the-art DP large-scale image classification, and the TopDown algorithm for the U.S. Decennial Census, observing that GDP fits the profiles remarkably well in all three cases. Although GDP is ideal for reporting the final guarantees, other formalisms (e.g., privacy loss random variables) are needed for accurate privacy accounting. We show that such intermediate representations can be efficiently converted to GDP with minimal loss in tightness.

$( varepsilon, δ)$ Considered Harmful: Best Practices for Reporting Differential Privacy Guarantees
Juan Felipe Gomez, Bogdan Kulynych, Georgios Kaissis, Jamie Hayes, Borja Balle, Antti Honkela
http://arxiv.org/abs/2503.10945

⏰ A quick reminder that Track I of the Health Privacy Challenge is still running strong, exploring privacy preservation in bulk RNA-seq datasets! We’re excited to see the innovative solutions Blue Teams and Red Teams will bring to the competition! #🫐🍅 #CAMDAConference #ISMB/ECCB2025

Exciting News! 🚀 We’re pleased to launch Track II of the Health Privacy Challenge 🫐🍅, focused on single-cell data! We invite the participants to explore the privacy and utility of synthetic single-cell RNA-seq data! #CAMDAConference
Register to participate: benchmarks.elsa-ai.eu?ch=4

Principal Investigator positions at ELLIS Institute Finland | ELLIS Institute Finland Now recruiting new PIs in artificial intelligence and machine learning

Call for Principal Investigators in #machinelearning & #artificialintelligence closes in 2 weeks! Why apply?
- Professorship affiliation at universities in Finland
- Research infrastructure incl. @lumi-supercomputer.eu
- Generous startup package
- @ellis.eu network
➡️ www.ellisinstitute.fi/PI-recruit

Principal Investigator positions at ELLIS Institute Finland | ELLIS Institute Finland Now recruiting new PIs in artificial intelligence and machine learning

I have big news: @ellis.eu has launched its 2nd major research center, @ellisfinland.bsky.social! I have agreed to start as founding director & the first call for PI positions is open. This is a major opportunity for outstanding researchers, join us! ellisinstitute.fi/PI-recruit

Helsinki Probabilistic Machine Learning Lab

1/ 🎉 We launched the *Helsinki Probabilistic Machine Learning Lab*, which combines multiple research groups at @univhelsinkics.bsky.social - and part of FCAI and ELLIS - working on, guess what, Probabilistic ML and AI.

And we are hiring! Please repost!

Website: www.helsinki.fi/probabilisti...

Less than a week left to apply. Come join us @csaalto.bsky.social or @univhelsinkics.bsky.social #postdocjobs #ai #artificialintelligence #datascience #postdoc #algorithms #cybersecurity

Very interesting challenge on privacy of synthetic data organised together with @steglelab.bsky.social and other ELSA project partners.

Please spread the word!

This would be rather suboptimal for those living further away from big centers or working in an area where most of the relevant community is elsewhere. Unfortunately we need common meetings to avoid fragmenting the community.

I have seen some anti-peer-review takes on here, and, believe me, I understand the frustrations, but I would urge caution before doing or publicly saying anything too drastic. We live in age rife with misinformation and anti-intellectualism—not everyone wants academia to survive

Risk is twofold: 1) Europeans risk losing access to models developed outside Europe, while 2) European researchers face significant additional burden to comply with the rules. This risks violating the freedom of research.

Open Letter on Regulation of Foundation Models in the AI Act and Research Open Letter on Regulation of Foundation Models in the AI Act and Research If you are a machine learning or AI researcher with at least a PhD and wish to sign the letter, please fill in this form: http...

Open letter: European AI researchers are worried that including research and open source in foundation model regulation in the AI Act could kill frontier AI research in Europe. docs.google.com/document/d/1...