Google DeepMind and the Government Technology Agency of Singapore have fine-tuned a scam detection model that runs entirely on a user’s phone, the two organizations said on Wednesday.
The problem is large and local. Scams make up more than half of Singapore’s reported crimes and have cost victims over S$4B since 2019. Social engineering now drives as much as 80 percent of those losses, as fraudsters move from crude phishing to long conversations that build trust before asking for money.
Why a generic model was not enough
Off-the-shelf models struggle with regional deception tactics. DeepMind took Google’s lightweight Gemma model and fine-tuned it on local scam conversations, covering impersonations of banks, public institutions, and well-known figures alongside trending schemes in the region.
The result classifies manipulative intent on the device rather than in a data center. When it detects a pattern, it raises a warning telling the user to pause before sharing credentials or money.
Privacy was the design constraint, not an afterthought. Because processing stays local, neither the Singapore government nor Google can read user conversations through the deployment. The model only reaches conversations with the user’s consent under the terms of the hosting application.
Battery life shaped the engineering. The team split detection into a two-stage pipeline so classification could run on modern smartphones without draining power.
The pair presented the work at Singapore International Cyber Week 2026 and framed it as a template for messaging platforms that must balance user protection against end-to-end encryption. Both sides said they would explore deploying similar technology in other countries.
