Andreas Terzis
@aterzis
Lead for GDM Security & Privacy Research team Past: Google Brain/Google/JHU CS Dept
Alex (alexbie98.github.io) and Umar(umarsyed.com) introduce an inference-only method for generating differentially private synthetic data. This approach is computationally cheaper than DP fine-tuning and can be extremely effective for quick prototyping.
Today we describe an inference-only approach to generating differentially private synthetic data via prompting off-the-shelf large language models with many examples in parallel and aggregating their responses in a privacy-preserving manner. Learn more at goo.gle/4izufJT
Join our Privacy & Security Research team at Google DeepMind to work on privacy-preserving technologies for Gemini and other GenAI-based agents
ML Privacy folks, we are looking for a person to join our team to work on contextually aware reasoning! Ping me if you have any questions! boards.greenhouse.io/deepmind/jobs/…
We recently updated the CaMeL paper, with results on new models (which improve utility a lot with zero changes!). Most importantly, we released code with it. Go have a look if you're curious to find out more details! Paper: arxiv.org/abs/2503.18813 Code: github.com/google-researc…
On the occasion of returning to Magenta's roots at @sonarplusd, we're dusting off the blog to share news and insights about what we're working on at @GoogleDeepMind on the Lyria Team. g.co/magenta/lyria-… Our latest post is about the Lyria RealTime API, providing access to…
We are starting our journey on making Gemini robust to prompt injections and in this paper we present the steps we have taken so far. A collective effort by the GDM Security & Privacy Research team spanning over > 1 year.
Our new @GoogleDeepMind paper, "Lessons from Defending Gemini Against Indirect Prompt Injections," details our framework for evaluating and improving robustness to prompt injection attacks.
Google presents Strong Membership Inference Attacks on Massive Datasets and (Moderately) Large Language Models
Gemma 3 explained: Longer context, image support, and a new 1B model. Read the blog for a deep dive into key enhancements → goo.gle/4lV8iaw
Excited to have @AnthropicAI and @GoogleDeepMind join @OpenAI in co-sponsoring our @AISecurityInst Agent Red-Teaming Challenge! New wave of challenges out tomorrow 1 pm EDT. Give it a try!
Major Update! The Agent Red-Teaming Challenge prize pool has surged from $130k to $170K. With @AnthropicAI & @GoogleDeepMind now co-sponsoring, the stakes have never been higher. This is the ultimate test for AI red teamers.
We are excited to participate in the agent red-teaming challenge to test the resilience of Gemini against prompt injection attacks! @GoogleDeepMind
Major Update! The Agent Red-Teaming Challenge prize pool has surged from $130k to $170K. With @AnthropicAI & @GoogleDeepMind now co-sponsoring, the stakes have never been higher. This is the ultimate test for AI red teamers.
With Adam Smith and @thejonullman, we have compiled a set of profiles of 29 people in the "foundations of responsible computing" community ("mathematical research in computation and society writ large") who are on the faculty job market. Check it out in the next tweet! 1/4 👇
This collaboration highlights the disconnect between expectations and current capabilities of unlearning and proposes ways for researchers and policymakers to close the gap.
Today we describe an inference-only approach to generating differentially private synthetic data via prompting off-the-shelf large language models with many examples in parallel and aggregating their responses in a privacy-preserving manner. Learn more at goo.gle/4izufJT
Excited to share that the Machine Learning and Optimization team at @GoogleDeepMind India is hiring Research Scientists and Research Engineers! If you're passionate about cutting-edge AI research and building efficient, elastic, and safe LLMs, we'd love to hear from you. Check…
We have an open position for a Research Scientist/Research Engineer to join our team! If interested: boards.greenhouse.io/deepmind/jobs/…