Google DeepMind and the Government Technology Agency of Singapore (GovTech) have announced an on-device scam detection model, developed under the National AI Partnership with the Singapore Government. The tool is designed to run on typical smartphones in Singapore and was presented at Singapore International Cyber Week (SICW 2026).
The announcement frames the effort as a response to a pressing safety challenge. Scams account for more than half of Singapore's reported crimes and have caused over S$4 billion in losses since 2019. Tactics have also shifted from simple phishing toward complex social engineering, in which attackers manipulate victims through convincing conversations before requesting money. Such scams now represent up to 80% of all losses.
To address localized deception, Google DeepMind fine-tuned its lightweight Gemma model on local scam conversations. These included impersonations of public institutions, banks, and local figures, as well as trending regional schemes, so the model can distinguish benign social chatter from genuine manipulation.
Crucially, the model processes conversation data on the device. Traditional threat detection relies on server-side scanning in the cloud, but many messaging platforms are end-to-end encrypted and cannot inspect private conversations. Bringing threat analysis to the device allows faster detection of manipulative intent and lets citizens act before financial or credential exposure occurs.
When manipulative conversational patterns are detected, the model triggers a real-time warning on the user's device advising them to pause and exercise caution before sharing sensitive information or funds. To manage battery efficiency, the system uses a two-stage pipeline: a lightweight classifier screens inbound messages with minimal power, and a fine-tuned Gemma model then evaluates a larger snapshot of the conversation to confirm a threat or discard a false alarm.
Governance measures included fine-tuning within GovTech's managed GCP environment under a "Sovereign Sandbox," excluding private or personally identifiable information and using anonymised datasets derived from scam reports and intelligence feeds. The model was stress-tested against synthetic dialogues and local corpora covering diverse scam categories and evolving impersonation tactics. The project suggests messaging platforms and operating system vendors could integrate such lightweight models to safeguard users locally without exposing private data.