- Google DeepMind and Singapore’s GovTech fine-tuned a Gemma model to detect conversational scams entirely on-device, with no cloud data egress.
- Scams account for more than half of Singapore’s reported crimes and over S$4 billion in losses since 2019, with social-engineering tactics driving up to 80% of that total.
- The system uses a two-stage pipeline: a lightweight classifier screens messages, then the fine-tuned Gemma model analyzes conversation context to confirm or dismiss a threat.
- Training used anonymized scam reports inside GovTech’s isolated cloud environment; neither Google nor the Singapore government can access users’ on-device conversations.
What Happened
Google DeepMind and Singapore’s Government Technology Agency (GovTech) fine-tuned Google’s lightweight Gemma model to detect conversational scams directly on a user’s device, the companies said October 14, 2026 on DeepMind’s blog, as part of their ongoing National AI Partnership. The work was presented at Singapore International Cyber Week 2026.
Why It Matters
Scams — investment fraud, fake job offers, impersonation of public officials — make up over half of Singapore’s reported crimes and more than S$4 billion in losses since 2019, with social-engineering tactics now accounting for up to 80% of that total, DeepMind said. Generic models often miss local deception patterns, which is why the system was fine-tuned on region-specific scam conversations.
Technical Details
The system runs a two-stage pipeline: a lightweight classifier screens inbound messages for potential threats using minimal power, then the fine-tuned Gemma model performs in-context analysis on a larger conversation snapshot to confirm or discard the alarm. All processing happens on-device, with no server-side message scanning — a different security posture than the cyberdefense focus of DeepMind’s Gemini 4 Argon launch.
Who’s Affected
GovTech Singapore’s Mark Chen, Director of Trust and Safety, said the partnership lets authorities “catch highly targeted, local scam attempts in real time.” Singapore citizens using apps that integrate the model get on-device warnings before sharing sensitive information, and neither Google nor the Singapore government can access the underlying conversations.
What’s Next
DeepMind said the rollout within apps will be phased and developed in partnership with the Singapore government, and the companies plan to explore deploying similar on-device scam detection in other countries.
