Google DeepMind and GovTech Singapore built a fine-tuned on-device scam detection model. It targets local conversational scams like impersonation and investment fraud. The model runs entirely on user devices with no cloud data access. A two-stage pipeline balances battery efficiency and accuracy. Warnings appear in real time when manipulative patterns are detected. Training used anonymised datasets from scam reports and intelligence feeds. Fine-tuning occurred in GovTech’s sovereign sandbox on GCP. The project was presented at Singapore International Cyber Week 2026. Rollout will be phased within applications in Singapore. Neither Google nor GovTech can access user conversations through deployment.
What changed
Scam detection relied on server-side cloud scanning of messages.
What it unlocks
Real-time on-device scam warnings without cloud data egress.
- S$4 billion in losses since 2019
- 80% of all scam losses from social engineering
- 2-stage on-device pipeline
What you need to act on it
- user consent
- application integration
- Singapore deployment
Sources