Google DeepMind and Singapore target scams with an on-device Gemma model

What is changing?

Google DeepMind announced that it is developing an on-device scam detection model with Singapore’s Government Technology Agency, GovTech, and the Ministry of Digital Development and Information. The lightweight Gemma-based system was fine-tuned for local scam conversations. Conversations are processed entirely on the phone, and deployment will allow access only with the user’s explicit consent.

The architecture works in two stages: a lightweight, fast classifier screens incoming messages, then passes suspicious conversations to a fine-tuned Gemma model that evaluates multi-turn context. The system produces either a scam alert or a benign verdict. This separation is intended to reserve the heavier analysis for conversations that appear risky.

Training and testing took place in GovTech’s sovereign sandbox on Google Cloud; according to DeepMind, no citizen personal data was used. The work relied on anonymized scam reports, threat intelligence and synthetic examples. The system is not yet generally available; phased deployment inside applications in Singapore is planned.