Google DeepMind and Singapore Government announce new AI tool to protect citizens against scams
The system fine-tunes Gemma to identify localized social engineering scams directly on smartphones without compromising end-to-end encryption.
- A two-stage pipeline pairs a lightweight fast screening classifier with a fine-tuned Gemma model for multi-turn contextual analysis.
- The model was fine-tuned on local scam patterns, including impersonations of Singaporean banks, public institutions, and local figures.
- Processing runs locally on the phone to deliver real-time warnings without sending private chat data to cloud servers.
- Training took place inside GovTech's isolated Google Cloud environment using synthetic dialogues and anonymized scam reports without citizen data.
Messaging platforms and operating systems can now detect social engineering fraud locally without breaking end-to-end encryption.
