The Pulse
Google DeepMind, Singapore Develop On-Device Scam Detection Model
Google DeepMind and Singapore’s Government Technology Agency have fine-tuned a Gemma model designed to detect scam conversations on devices. Its rollout within applications is planned in phases, with access to conversations subject to user
AI.info Team · Written with AI ·
Google DeepMind and Singapore’s Government Technology Agency, or GovTech, have fine-tuned a Gemma model to detect scam conversations on users’ devices, an effort the partners say can flag suspicious messages without exposing private chats to cloud servers. DeepMind announced the collaboration on October 14, 2026, and said it was presented at Singapore International Cyber Week 2026.
Gemma checks suspicious conversations on the phone
The system uses two stages to limit the work required from a smartphone. A lightweight classifier first screens incoming messages for possible threats; only messages it flags go to the fine-tuned Gemma model for a broader analysis of conversation context. That second check can confirm a scam or dismiss a false alarm, according to DeepMind.
If the model detects manipulative patterns, it can issue a warning on the device telling the user to pause before sharing money or sensitive information. DeepMind says the model was trained to recognize local tactics, including impersonations of public institutions, banks and local figures, as well as regional schemes. The company describes the goal as distinguishing ordinary conversation from attempts to manipulate a victim.
“Fighting scams shouldn't mean sacrificing personal privacy. Working with Google DeepMind, we're bringing advanced AI protection directly onto citizens' devices. This allows us to catch highly targeted, local scam attempts in real time while keeping personal communications private.”
Mark Chen, Director, Trust and Safety, GovTech Singapore
Private messages stay on the device
Running the analysis locally addresses a problem for services that use end-to-end encryption: messages are not available for conventional server-side scanning. DeepMind says the model will process conversation data on the device, and that neither the Singapore Government nor Google will be able to access users’ conversations through its deployment. The model will only access on-device conversations with the user’s consent and under the terms of the application that integrates it.
That consent condition makes the app’s role central. The announcement does not name a messaging service or app that will use the model, and it does not give a public release date. Instead, DeepMind says deployment within applications will be phased and developed with the Singapore Government.
Training data and performance details remain limited
DeepMind says fine-tuning took place in a GovTech-managed Google Cloud environment, which it calls a “Sovereign Sandbox.” The company says the training excluded private and personally identifiable citizen data, using anonymised datasets derived from scam reports and intelligence feeds. It also says the model was tested with synthetic conversations and local datasets covering multiple scam categories.
The announcement offers no detection-accuracy results, false-positive rate, model size or battery measurements. Those details would help users and prospective app partners assess how the system performs in practice, particularly when scam tactics change or conversations include harmless discussions of money and personal information.
Singapore’s scam figures frame the project
DeepMind says scams account for more than half of Singapore’s reported crimes and have caused over S$4 billion in losses since 2019. It also says social-engineering scams, in which perpetrators build trust through convincing conversations before asking for money, account for up to 80% of losses.
The collaboration’s next concrete step is integration into applications, but the announcement names no participating app or timetable. For now, it describes a locally processed detection model, a two-stage screening design and a phased rollout to be developed with the Singapore Government.