DEV Community

Cover image for Gboard Privacy Myth: Your Texts Still Leak via Updates
Tom Morgan
Tom Morgan

Posted on

Gboard Privacy Myth: Your Texts Still Leak via Updates

Their attack targets individual updates; central differential privacy protects the aggregate model.

The claims are Google’s. Frequently asked question: Is federated learning actually private? The Dublin attack used a public text-message corpus, an extracted model, and an assumed curious server.

Key takeaways

  • Sources: Suliman & Leith (2022), Two Models Are Better than One: Federated Learning Is Not Private for Google GBoard Next Word Prediction Xu et al.
  • (2023), Federated Learning of Gboard Language Models with Differential Privacy, Zhang et al.
  • (2023), Private Federated Learning in Gboard McMahan & Thakurta (2022), Federated Learning with Formal Differential Privacy Guarantees Xu & Zhang…

(2024), Federated Learning in Practice: Reflections and Projections (Daly & Ramage, 2026), Toward provably private learning from federated data, and the paper Hard et al.

Read the full article: https://www.aipersonalization.cloud/gboard-privacy-myth/

Top comments (0)