Google DeepMind and Singapore's GovTech agency have fine-tuned a lightweight Gemma model to recognise local scams and run it entirely on the phone. The project was shown on 14 October at Singapore International Cyber Week, as part of the national AI partnership with the Singapore Government.
Key takeaways
- Scams make up over half of Singapore's reported crimes and more than S$4 billion in losses since 2019
- Social engineering now accounts for up to 80% of all losses
- The model analyses conversations locally — neither the Singapore Government nor Google can access them
- A two-stage pipeline: a lightweight classifier screens messages, then the fine-tuned Gemma confirms the threat
- Fine-tuning ran inside a GovTech-managed cloud environment, with no citizens' personal data
Why a generic model was not enough
Generic models struggle with local deception techniques, so Google DeepMind fine-tuned the lightweight Gemma on sets of local scam conversations — from impersonations of public institutions, banks and recognisable figures to regional schemes currently in circulation.
The model has to tell benign chatter from genuine social engineering once it is built into apps citizens install.
Two stages, so the battery survives
Running a complex classifier on a modern smartphone is hard mainly because of power draw. Hence the two-stage pipeline: first a lightweight, efficient classifier screens inbound messages and catches potential threats, and only then does the fine-tuned Gemma analyse a larger snapshot of the conversation to confirm the threat or discard a false alarm.
The diagram reflects the order of steps described by the authors. It does not show the model's architecture or how it integrates with a specific app.
When the model spots a manipulative pattern, a real-time warning appears on the device — before the user hands over credentials or transfers money.
Privacy as a precondition
Fighting scams shouldn't mean sacrificing personal privacy. Working with Google DeepMind, we're bringing advanced AI protection directly onto citizens' devices.
Mark Chen, Director, Trust and Safety, GovTech Singapore.
Unlike classic server-side message scanning, conversation data never leaves the phone — this is on-device AI acting as a safeguard. The model reaches that data only with the user's consent, under the terms of the specific app.
| Rule | What it means in practice |
|---|---|
| Sovereign sandbox | fine-tuning inside a GovTech-managed cloud environment, training data stays put |
| No citizen data | training on anonymised sets derived from scam reports, no personal data |
| Stress-testing | the model checked against synthetic dialogues and local corpora |
Why it matters
Any messenger with end-to-end encryption has faced a choice: protect the content, or scan it for fraud. Singapore shows a third route and, with it, a template in which a government contributes local threat context and the model provider contributes weights and engineering. The open question is efficacy: neither party published a detection rate or a false-alarm figure.
What's next
- The rollout inside apps will be phased, run together with the Singapore Government
- Google says it is exploring similar deployments in other countries
