Google DeepMind and Singapore’s GovTech developed a smartphone-based scam detector using a locally fine-tuned Gemma model to flag manipulative conversations.
Google DeepMind and Singapore’s Government Technology Agency (GovTech) have developed an on-device scam detection model designed to identify manipulative conversations while keeping message analysis on the user’s smartphone. Announced October 14, the project adapts a lightweight Gemma model to scam patterns found in Singapore, including impersonations of public institutions, banks and local figures, as well as regional schemes.
The system uses a two-stage pipeline. An initial step screens conversations for suspicious patterns; when it detects possible manipulation, a fine-tuned Gemma model analyzes the conversation across multiple turns and returns either a scam alert or a benign verdict. The tool can then issue a real-time warning advising the person to pause before sharing sensitive information or sending money.
The announcement addresses a privacy challenge for messaging services: many use end-to-end encryption, which prevents providers from inspecting message contents on their servers. By performing analysis on the device, the model is intended to identify conversational warning signs without sending private chats to a cloud-based detection system. DeepMind says the model was optimized for typical smartphones in Singapore, where battery efficiency is an important consideration.
The project was developed through a national AI partnership between Google DeepMind and the Singapore government. DeepMind says fine-tuning took place in a GovTech-managed Google Cloud environment, using anonymized data derived from scam reports and intelligence feeds rather than private or personally identifiable information. The model was also tested with synthetic dialogues and local scam-related material.
The partners presented the work at Singapore International Cyber Week 2026. The announcement describes a model that messaging platforms and operating-system vendors could integrate; it does not specify a public rollout date or announce a consumer deployment. The approach illustrates how locally adapted models may support fraud warnings while leaving message contents on users’ devices.