Google Research Moves Federated Learning Into TEEs: Gboard Now Trains With Externally Verifiable Differential Privacy
Google Research has announced a next-generation Federated Learning (FL) system built on Trusted Execution Environments (TEEs).The research team claims externally verifiable central differential privacy (DP) guarantees for FL for the first time.What Problem Does TEE-Based Federated Learning Solve?
Google introduced Federated Learning FL in 2017.It powers next-word prediction and Smart Compose on Gboard, reply suggestions in Google Messages, and Smart Text Selection in Android.Earlier systems had a trust gap.
Devices uploaded data for immediate aggregation, but outsiders could not verify that data was never logged or inspected.Secure Aggregation added cryptographic protection.However, it was not compatible with state-of-the-art central DP algorithms like matrix factorization DP-FTRL.
Google also had to be trusted to add DP noise correctly.The new design moves client gradient computation to the server.It then makes that server logic attestable, so the operator no longer needs to be trusted.How Does the System Work?
The system builds on Google’s earlier confidential federated analytics work.It coordinates 4 core components: Data upload: Devices encrypt training examples locally and pre-authorize an access policy.The policy lists which TEE computations may process the data.
Policies must appear in a public transparency log.KMS and policy verification: A Key Management System, built from TEEs running the RAFT consensus protocol, releases keys only to workloads matching the policy.
Workload execution: A root TEE runs a Python training loop and delegates subtasks to worker TEEs.Orchestration uses Federated Language, derived from TensorFlow Federated.Only DP model weights are released.
Fault-tolerant recovery: Each round saves a KMS-encrypted recovery state for handling root or worker failures.Why Is the Privacy Guarantee Verifiable?Access policies are published to Rekor, Sigstore’s public transparency log.External auditors can track every server workload a device could feed.
The KMS and data processing binaries are reproducibly buildable from open source code.The policies directly describe the Python training program.To protect proprietary model architectures, TEEs support sideloading serialized logic at runtime.
All privacy-relevant logic must stay hardcoded in the attested program.Workload operators see only metrics and DP model weights.Encrypted data can be decrypted only for a limited time after upload.What Did Gboard Gain?
Gboard used the system to launch English and Japanese next-word prediction models with stronger privacy guarantees and improved accuracy.Two design choices drive this: First, all uploads are collected before server-side training runs.Diurnal swings in device availability no longer slow training.
The program can compute an optimal participation schedule and tune DP parameters.Google’s privacy-utility curves come from training an English model for 5000 rounds with cohorts of 6500 devices on both systems.Second, the bottleneck moved to the server.
Previous FL models took 1 to 2 months each to train.Training now parallelizes across machines, limited only by TEE resource availability.Google reports substantially faster compute times but does not publish a single speedup figure.
Interactive Explainer: Inside the TEE-Based FL Pipeline #mtp-tfl{--b:#4285F4;--r:#EA4335;--y:#FBBC05;--g:#34A853;--ink:#202124;--mut:#5f6368;--line:#dadce0;--soft:#f8f9fa;font-family:"Google Sans",Roboto,Arial,sans-serif!important;color:var(--ink)!important;background:#ffffff!
important;border:1px solid var(--line)!important;border-radius:16px!important;padding:22px!important;max-width:860px;margin:0 auto;box-sizing:border-box;line-height:1.5} #mtp-tfl *{box-sizing:border-box} #mtp-tfl .
bar{display:flex;height:4px;border-radius:4px;overflow:hidden;margin-bottom:16px} #mtp-tfl .bar i{flex:1} #mtp-tfl h3{margin:0 0 4px!important;font-size:21px!important;color:var(--ink)!important;font-weight:700} #mtp-tfl .sub{margin:0 0 16px!important;color:var(--mut)!
important;font-size:14px} #mtp-tfl .tabs{display:flex;gap:8px;flex-wrap:wrap;margin-bottom:16px} #mtp-tfl .tab{border:1px solid var(--line)!important;background:#fff!important;color:var(--ink)!
important;border-radius:999px;padding:8px 16px;font-size:14px;cursor:pointer;font-family:inherit;transition:all .2s} #mtp-tfl .tab.on{background:var(--b)!important;color:#fff!important;border-color:var(--b)!important} #mtp-tfl .pane{display:none} #mtp-tfl .pane.on{display:block;animation:mtpfade .
35s ease} @keyframes mtpfade{from{opacity:0;transform:translateY(6px)}to{opacity:1;transform:none}} #mtp-tfl .stage{background:var(--soft)!important;border:1px solid var(--line)!important;border-radius:12px;padding:10px} #mtp-tfl svg{width:100%;height:auto;display:block} #mtp-tfl .
node rect{fill:#fff;stroke:var(--line);stroke-width:2;transition:all .3s} #mtp-tfl .node text{font-size:13px;fill:var(--ink);font-family:inherit;text-anchor:middle} #mtp-tfl .node .s{font-size:10.5px;fill:var(--mut)} #mtp-tfl .node.act rect{stroke-width:3} #mtp-tfl .
path{fill:none;stroke:var(--line);stroke-width:3;stroke-dasharray:6 6} #mtp-tfl .path.act{stroke-dasharray:6 6;animation:mtpdash 1s linear infinite} @keyframes mtpdash{to{stroke-dashoffset:-24}} #mtp-tfl .pkt{opacity:0} #mtp-tfl .pkt.act{opacity:1} #mtp-tfl .
steps{display:flex;gap:6px;margin:14px 0 10px;flex-wrap:wrap} #mtp-tfl .sb{flex:1;min-width:120px;border:1px solid var(--line)!important;background:#fff!important;color:var(--ink)!
important;border-radius:10px;padding:8px 10px;font-size:13px;cursor:pointer;text-align:left;font-family:inherit;transition:all .2s} #mtp-tfl .sb b{display:block;font-size:11px;letter-spacing:.06em;text-transform:uppercase} #mtp-tfl .sb.on{color:#fff!important} #mtp-tfl .
card{border-left:4px solid var(--b);background:#fff!important;border-radius:8px;padding:12px 14px;font-size:14.5px;color:var(--ink)!important;border-top:1px solid var(--line);border-right:1px solid var(--line);border-bottom:1px solid var(--line)} #mtp-tfl .
card strong{display:block;margin-bottom:4px;font-size:15px} #mtp-tfl .ctl{display:flex;gap:8px;margin-top:10px;flex-wrap:wrap} #mtp-tfl .btn{border:none;background:var(--b)!important;color:#fff!important;border-radius:8px;padding:8px 14px;font-size:13.
5px;cursor:pointer;font-family:inherit} #mtp-tfl .btn.ghost{background:#fff!important;color:var(--b)!important;border:1px solid var(--b)!important} #mtp-tfl .sw{display:inline-flex;border:1px solid var(--line);border-radius:999px;overflow:hidden;margin-bottom:14px} #mtp-tfl .
sw button{border:none;background:#fff!important;color:var(--ink)!important;padding:8px 16px;font-size:13.5px;cursor:pointer;font-family:inherit} #mtp-tfl .sw button.on{background:var(--ink)!important;color:#fff!important} #mtp-tfl .
grid{display:grid;grid-template-columns:repeat(2,1fr);gap:10px} #mtp-tfl .cell{border:1px solid var(--line);border-radius:10px;padding:12px;background:#fff!important;transition:all .35s} #mtp-tfl .cell .k{font-size:12px;color:var(--mut);text-transform:uppercase;letter-spacing:.05em} #mtp-tfl .cell .
v{font-size:15px;font-weight:600;margin-top:4px} #mtp-tfl .cell.ok{border-color:var(--g);background:#e6f4ea!important} #mtp-tfl .cell.ok .v{color:#137333} #mtp-tfl .cell.no{border-color:var(--y);background:#fef7e0!important} #mtp-tfl .cell.no .v{color:#b06000} #mtp-tfl .
note{font-size:13px;color:var(--mut);margin-top:10px} #mtp-tfl .opts{display:grid;grid-template-columns:1fr 1fr;gap:10px;margin-bottom:12px} #mtp-tfl .opt{border:2px solid var(--line);border-radius:12px;padding:12px;cursor:pointer;background:#fff!important;font-size:14px;transition:all .
2s} #mtp-tfl .opt code{display:block;font-family:"Roboto Mono",monospace;font-size:11.5px;background:var(--soft)!important;color:var(--ink)!important;padding:6px;border-radius:6px;margin-top:6px;word-break:break-all} #mtp-tfl .opt.on{border-color:var(--b)} #mtp-tfl .
kms{display:flex;align-items:center;gap:12px;padding:12px;border-radius
Related
相關文章
Announcing Project ProtoStar: Making Publicly Funded Science AI-Ready
Back to Articles Announcing Project ProtoStar: Making Publicly Funded Science AI-Ready Published October 29, 2026 Update on GitHub Upvote - Avijit Ghosh evijit Follow hfpolicy Sayeed Choudhury sayeedmd Follow CMU-OSPO Greg Tananbaum gtananbaum Follow ORCA-Open Emma Scharfmann EmmaScharfmann Follow hugging-science Georgia Channing cgeorgiaw Follow hugging-science Eunice Mercado-Lara EuniceML Follow ORCA-Open Irene Solaiman irenesolaiman Follow hfpolicy An initiative bringing together universities, research organizations, industry, and the open science community to make publicly funded research easier to discover, reuse, and build upon.
Build specialized multimodal models with LFM3
Back to Articles Build specialized multimodal models with LFM3 Team Article Published October 27, 2026 Upvote 11 +5 Leonie Monigatti iamleonie Follow LiquidAI Jimmy T.
上新風向標，抖音商城雙 11 年度新品必買榜出爐
雙11大促正酣，抖音商城發佈2026年度新品必買榜TOP100，覆蓋護膚、彩香、鞋靴、耐消等類目。榜單聚焦具有引領性的新品賽道，精選最具代表性的100款新品，讓每一件好物都被看見。從細胞級抗老的成分進階，到一盒多用、一抹成妝的美妝新風；從城市與山野之間的足下美學，到懂生活懂情緒的治癒系家電產品。這份年度新品必買榜，集合了當下最前沿的消費趨勢與生活方式提案，無論你嚮往哪一種新生活，都能在這裡找到答案。01 護膚：功效進階，進入「成分效度」時代護膚榜的四條主線，勾勒出「功效進階」的完整鏈路。
Ψ-0.6把感知語言幾何擰進統一模型
Computer Science > Artificial Intelligence arXiv:2610.32177 (cs) [Submitted on 23 Oct 2026] Title:$Ψ$-0.
Block發佈全雙工語音模型Dessa
AINewsJack Dorsey’s Block made an AI model that talks to you like an awkward new friendYou have got to listen to these conversations with Block’s new Dessa AI model.
Kyutai CADRE: Video Autoencoders That Cut World Model Lag Below 230 ms
Kyutai has released CADRE, open causal diffusion representation autoencoders for interactive world models. The Kyutai CADRE video autoencoder targets one problem: action lag. Many world models borrow autoencoders from text-to-video generators.