Google DeepMind and Singapore build an on-device Gemma model to catch scam messages

Google DeepMind and Singapore's Government Technology Agency have unveiled a scam-detection model that runs entirely on a user's phone. Built under the two organisations' National AI Partnership and presented at Singapore International Cyber Week, the system fine-tunes Google's compact Gemma model on examples of local scam conversations, including impersonations of public institutions, banks and well-known figures.

Privacy is designed in rather than bolted on. The model analyses conversations on the device, so neither the Singapore government nor Google can see users' messages. It accesses conversations only with the user's consent under the host app's terms, and fine-tuning took place in GovTech's managed cloud environment using data stripped of private or personally identifiable information.

The architecture is pragmatic. A lightweight classifier screens every incoming message first; only those it flags as suspicious are passed to the fine-tuned Gemma model, which performs multi-turn analysis to confirm a threat or dismiss a false alarm. That keeps battery and latency costs low while reserving the heavier model for cases that need it.

GovTech has launched an open beta of the resulting Anti-Scam Pango app for Android users in Singapore, giving the model a real-world deployment from day one.

Why it matters

Scam losses are a mass-market harm, and on-device inference is the only approach that scales to a whole population without creating a surveillance system. The project demonstrates that a small open model, fine-tuned on local data, can deliver a public-safety outcome that a frontier model in the cloud cannot, for reasons of privacy and cost.

Indicops view

India faces the same problem at far larger scale, with digital-arrest and UPI fraud draining billions of rupees a year. The Singapore design is directly transferable: an open model fine-tuned on Indian-language scam patterns, running on-device, distributed through banks and telecom operators. We see this as one of the highest-value public AI projects a state or central agency could commission, and the model and tooling to do it already exist.