Google DeepMind and Singapore government unveil two-stage on-device AI fraud detector

The system, revealed at Singapore International Cyber Week 2026, uses a two-stage architecture. A lightweight classifier scans incoming messages with minimal power draw and flags potential threats. Only borderline cases reach the second stage, a fine-tuned Gemma model that examines broader conversation context to confirm or dismiss the suspicion. Most messages are rejected at the first layer.

Development took place under a strict governance framework. Fine-tuning ran inside a "sovereign sandbox" — a managed GCP environment operated by GovTech, the government technology agency — that keeps training data isolated. The teams say no personal or identifying citizen data was used; training data came from anonymous fraud reports and intelligence feeds. The model was evaluated against synthetic dialogues and local corpora to cover a wide range of impersonation scenarios and evolving tactics.

The practical upshot: messaging platforms and operating-system vendors can embed these lightweight models directly on the device, without uploading conversation content to the cloud or breaking end-to-end encryption. The project demonstrates that an active, on-device protection layer can run locally while preserving user privacy — a government–model-provider collaboration model built for public benefit.