DeepMind and GovTech Singapore ship on-device Gemma scam filter

Google DeepMind and Singapore's Government Technology Agency have shipped a scam-detection model that runs entirely on a user's phone, trained inside a government-managed cloud enclave the two sides are calling a 'Sovereign Sandbox.' The pitch, in DeepMind's announcement, is that neither Google nor the Singapore government can read the messages the detector screens.

The problem on paper is large. Scams comprise over 50% of Singapore's reported crimes, total losses have exceeded S$4 billion since 2019, and complex social engineering schemes now account for up to 80% of financial losses, according to the post. The system is built as a two-stage pipeline. A lightweight classifier screens incoming messages with minimal power consumption, and anything it flags is handed to a fine-tuned Gemma that analyzes broader conversation context to confirm the threat or drop the false positive.

Fine-tuning happened inside GovTech Singapore's managed GCP environment on anonymized scam reports covering impersonations of public institutions, banks and regional schemes, with stress-testing against synthetic dialogues and local scam patterns. 'Fighting scams shouldn't mean sacrificing personal privacy,' said Mark Chen, Director of Trust and Safety at GovTech Singapore. 'Working with Google DeepMind, we're bringing advanced AI protection directly onto citizens' devices.'

The post publishes no accuracy figures, no false-positive rate, and does not say which Gemma variant ships or how large the on-device footprint is. It lands in a visible run of edge-AI launches we have been tracking in our edge-AI feed, where frontier labs keep pushing inference off the server and onto the handset. What is new here is the governance wrapper: a national AI agency fine-tuning inside its own cloud sandbox, with the vendor cut out of the inference path.