DeepMind and Singapore Build On-Device Scam Detector

Google DeepMind and Singapore's Ministry of Digital Development and Information (MDDI) have built an on-device scam detection system based on a fine-tuned Gemma model. The work was presented at Singapore International Cyber Week (SICW 2026).

The system works in two stages. First, a small, low-power classifier screens incoming messages for possible threats. If one is flagged, the fine-tuned Gemma model reviews a larger part of the conversation and either confirms the threat or drops it as a false alarm. Everything runs on the device, so messages are never sent off the phone for analysis. DeepMind says this makes the approach compatible with end-to-end encrypted messaging.

According to DeepMind, fine-tuning was done in a "sovereign sandbox": a Google Cloud environment run by GovTech, Singapore's government technology agency, which kept the training data isolated. The training data left out private and personally identifiable information. Instead it used anonymised datasets drawn from scam reports and intelligence feeds. The model was tested against synthetic dialogues and local text collections covering a range of scam types and impersonation tactics.

DeepMind presents the project as a template for messaging platforms and operating system vendors, which could build these lightweight models directly into their products. It also casts it as an example of governments and model providers working together.

The announcement does not include accuracy, false-positive rates, model sizes or latency figures. It also does not name any messaging apps or device makers that have agreed to deploy the system, or say when it will reach users.