media MarkTechPost · 1h ago · releases

Google Research moves Federated Learning into TEEs for externally verifiable differential privacy

from English

Google Research has announced a next-generation Federated Learning system built on Trusted Execution Environments (TEEs) that provides externally verifiable central differential privacy guarantees. This architecture shifts client gradient computation to the server, allowing the server logic to be attested so operators no longer need to be trusted with raw data.

Devices encrypt training examples locally and pre-authorize access policies published in Sigstore’s Rekor transparency log.
A Key Management System using the RAFT consensus protocol releases keys only to workloads matching the policy.
A root TEE runs a Python training loop via Federated Language, delegating subtasks to worker TEEs while releasing only DP model weights.
Gboard launched English and Japanese next-word prediction models with stronger privacy guarantees and improved accuracy.
Training time is reduced from 1-2 months per model by parallelizing across machines, limited only by TEE resource availability.

The system allows external auditors to track every server workload a device could feed, ensuring that privacy-relevant logic remains hardcoded in the attested program.

Importance 2/3 New feature vs. leaders MarkTechPost Google DeepMind Safety & alignment Training methods

Related articles

arxiv arXiv cs.LG · 47d ago · 35 views

Debate Training Reduces Reward Hacking in RLAIF

Researchers demonstrate that using debate—a two-player adversarial game between a generator and a critic adjudicated by a weaker LLM judge—reduces reward hacking compared to standard reinforcement learning from AI feedback (RLAIF). In this setup, the policy learns to exploit systematic errors in its AI judge, degrading task performance, but debate maintains judge performance throughout training.

arxiv arXiv cs.CL · 107d ago · 15 views

Sequential DPO Shows Variable Preference Impact Across Settings

A study of sequential Direct Preference Optimization finds that later training does not uniformly degrade earlier learned preferences. The effect varies by objective relationship, signal strength, and training order, ranging from partial degradation to positive transfer. Pair-level analysis reveals heterogeneous changes, with high-confidence preference pairs sometimes improving despite aggregate metric stability.

media TLDR AI · 3d ago · 14 views

Google announces Gemini 4 Argon; NVIDIA releases OpenShell for autonomous agents

Google has announced Gemini 4 Argon, a frontier model designed for sustained reasoning in software engineering, enterprise knowledge work, and cybersecurity. Access is initially limited to trusted cyber defenders, with broader availability planned after phased safety testing.

lab Google DeepMind Blog · 4d ago · 36 views

Google DeepMind introduces SynthID Bio to watermark AI-generated proteins

Google DeepMind has introduced SynthID Bio, a proof-of-concept technology that embeds imperceptible watermarks directly into the biological code of AI-designed proteins and structures. This approach ensures the signature is verifiable on the synthesized physical protein itself while preserving its biological function.

arxiv arXiv cs.CL · 5d ago · 14 views

Study reveals epistemic policy divergence in multi-turn LLM contamination

A study titled "Epistemic Policy Divergence in Multi-Turn LLM Contamination: A Protocol-Gradient Investigation" evaluates how large language models handle false premises injected into conversation history, a failure mode termed session-level contamination. The researchers tested GPT-5.4 Mini, Gemini-3.1 Flash-Lite, and GLM-4.5-Air across ten knowledge domains using 22,500 turns at temperature zero.