Google DeepMind and Singapore’s GovTech announced a scam detection model designed to examine conversations on the user’s phone. A lightweight screening stage calls on a tuned Gemma model when a message looks suspicious. The partners plan phased deployment in applications, with user consent and local conversation processing built into the design.
Google DeepMind and GovTech, Singapore’s government technology agency, announced an on-device scam detection model on 14 October.[1], [2]
The project brings fraud detection onto the user’s phone. Developed under their National AI Partnership, it was presented during Singapore International Cyber Week, with deployment in applications planned in phases.[1]
An efficient classifier first checks incoming messages for possible threats. Potential threats are passed to a tuned Gemma language model, which examines a larger conversation snapshot for manipulation or a false alarm. DeepMind says this two-stage approach is intended to limit battery and computing demands. When a pattern is detected, the device displays a warning before the user shares information or money. Training examples include impersonation of public institutions, banks and local figures.[1]
DeepMind says conversation analysis runs locally and requires user consent under the host application’s terms. Neither Google nor the Singapore government can access conversations through this deployment, according to the announcement. Training used a GovTech-managed Google Cloud environment and anonymised information derived from scam reports and intelligence feeds. Testing included synthetic dialogues and local text collections covering different fraud categories and changing impersonation tactics.[1]
