Google DeepMind and Singapore build on-device Gemma model to detect scams

Google DeepMind announced on October 14, 2026 that it has worked with Singapore's Government Technology Agency (GovTech Singapore) to create an AI model that spots scams as they unfold in conversations. Built by fine-tuning Google's lightweight Gemma model, the system analyses messages locally on the device rather than in the cloud. The collaboration was presented at Singapore International Cyber Week (SICW 2026).

According to Google DeepMind, scams such as investment fraud, bogus job offers and impersonation of officials make up more than half of crimes reported in Singapore and have cost victims over S$4 billion since 2019. The company says criminals have moved beyond basic phishing toward social engineering, building trust through persuasive chats before asking for money, and that these schemes now drive as much as 80% of losses.

Trained on local tactics

Because general-purpose models often miss region-specific deception, the team trained Gemma on local scam dialogues, including impersonations of public bodies, banks and well-known local figures, along with emerging regional schemes. The goal is to tell ordinary chatter apart from genuine manipulation. When the model detects a suspicious pattern, it shows a real-time warning on the phone urging the user to stop before handing over money or sensitive details.

To keep battery drain manageable, the system uses two stages. A small, low-power classifier first screens incoming messages for possible threats. Only then does the fine-tuned Gemma model examine a wider slice of the conversation to confirm the danger or dismiss a false alarm.

Privacy and governance

Google DeepMind says the model will access on-device conversations only with user consent and under the terms of whichever app hosts it. Since processing stays on the phone, the company states that neither Google nor the Singapore Government can read users' chats through the deployment. Fine-tuning took place in a GovTech-managed Google Cloud environment, used anonymised data drawn from scam reports and intelligence feeds with no personal information, and was stress-tested against synthetic dialogues and local text collections.

Rollout inside apps will be phased and coordinated with the Singapore Government, with no timeline given. The work sits under the two parties' National AI Partnership, and Google DeepMind said it will explore whether similar technology could be used in other countries.

The project is pitched as a template for encrypted messaging services that want to protect users without scanning messages on servers. On-device analysis promises speed and stronger privacy, though how well the model performs at scale will only become clear once it reaches real users.