A new on-device scam detection model developed with the Government Technology Agency of Singapore is designed to spot manipulative conversations without sending users’ chats to cloud servers. The system uses local scam patterns to identify threats and warn users before they share sensitive information or money.
Key takeaways
The model combines a lightweight message classifier with a fine-tuned Gemma model.
Conversation access requires user consent, and processing stays on the device.
Singapore has recorded more than S$4 billion in scam losses since 2019.
Deployment within applications will proceed in phases with the Singapore Government.
Google DeepMind said in an October 14, 2026, blog post that it developed the model with the Government Technology Agency of Singapore , known as GovTech Singapore. The collaboration forms part of its National AI Partnership with the Singapore Government.
Singapore partnership unveiled at SICW 2026
The project was presented at Singapore International Cyber Week (SICW 2026) . The company described it as a model for messaging platforms seeking to protect users while maintaining end-to-end encryption.
According to the blog post, scams account for more than half of Singapore’s reported crimes and have caused more than S$4 billion in losses since 2019 . Examples include investment fraud, fake job offers and impersonation of public officials.
The post describes a shift from simple phishing toward convincing conversations that manipulate victims before requesting money. These social engineering scams account for up to 80% of all losses .
How on-device scam detection uses local context
The system uses a two-stage detection pipeline : a low-power classifier screens incoming messages, then a fine-tuned Gemma model examines the broader conversation. The second stage assesses context to confirm a threat or discard a false alarm.
The company fine-tuned Google’s lightweight Gemma model on local scam conversations, including impersonations of public institutions, banks and local figures, alongside trending regional schemes. The approach addresses local deception techniques that generic models often struggle to recognise.
The first-stage classifier uses minimal power to identify potential threats. The more detailed conversation analysis follows in the second stage, addressing the battery-efficiency challenge of running a complex scam classifier on smartphones.
When deployed in applications, the model will access conversations only with user consent , under the application’s terms of use. Processing runs entirely on the device; neither the Singapore Government nor the company will gain access to users’ conversations through the deployment.
Training inside GovTech Singapore’s environment
The fine-tuning process occurred inside a GCP environment that GovTech Singapore manages and maintains, ensuring the training data remained isolated. Training excluded private and personally identifiable information, using anonymised datasets derived from scam reports and intelligence feeds.
The team also tested the model against synthetic dialogues and local corpora covering different scam categories and evolving impersonation tactics.
Real-time warnings before users share funds
When the model identifies manipulative conversational patterns, it issues a real-time warning on the user’s device. The alert urges users to stop and think carefully before handing over sensitive information or money.
Mark Chen, Director, Trust and Safety at GovTech Singapore, said: “This allows us to catch highly targeted, local scam attempts in real time while keeping personal communications private.”
The rollout of on-device scam detection within applications will be phased and developed with the Singapore Government.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
