Google DeepMind and Singapore's Government Technology Agency (GovTech) have unveiled an on-device scam detection model designed to warn smartphone users about manipulative conversations in real time — without private messages ever leaving the phone. The announcement, published on the Google DeepMind blog and presented at Singapore International Cyber Week 2026, describes a lightweight model fine-tuned for the local scam landscape and optimized to run on typical smartphones.
Why Singapore Is the Testing Ground For more context on this story, see our ongoing AI news.
The scale of the scam problem in Singapore gives the project a clear rationale. According to the announcement, authored by Google DeepMind's JJ Geewax and Miguel de Andrés, scams make up more than half of the country's reported crimes and have caused over S$4 billion in losses since 2019. Scam tactics, the companies say, have shifted from simple phishing toward complex social engineering, in which attackers build convincing conversations over days or weeks before asking for money — a pattern that now accounts for up to 80% of all scam losses.
Generic AI models often struggle with localized deception techniques, the announcement notes. To address that, the team fine-tuned Google's lightweight Gemma model on local scam conversations, ranging from impersonations of public institutions, banks, and local figures to trending regional schemes. That training is intended to let the model distinguish harmless social chatter from genuine manipulation.
How the Two-Stage On-Device Pipeline Works
The model processes conversation data entirely on the device itself. That design choice matters because conventional threat detection scans messages on cloud servers — something that is impossible for the many messaging platforms that use end-to-end encryption, where providers cannot inspect private conversations even if they wanted to.
Running the analysis locally also makes detection faster, so people can act before financial or credential exposure occurs. When the model detects manipulative conversational patterns, a real-time warning appears on the device urging the user to pause before sharing sensitive information or money.
Battery efficiency was a critical concern in running a complex scam classifier on modern smartphones, according to the announcement. The teams solved this with a two-stage pipeline:
- Stage 1: A lightweight, efficient classifier screens inbound messages using minimal power, focused on catching potential threats.
- Stage 2: The fine-tuned Gemma model performs a contextual evaluation of a larger snapshot of the conversation to confirm a threat or discard a false alarm.
Governance: No Citizen Data, Sovereign Sandbox
The collaboration was presented at Singapore International Cyber Week 2026, whose opening address was delivered on October 13 by K Shanmugam, Singapore's Senior Minister and Minister for Home Affairs, according to the Ministry of Home Affairs. The announcement describes strict governance maintained throughout development under three measures it labels Sovereign Sandbox, No Citizen Data, and Stress-Testing.
Under Sovereign Sandbox, fine-tuning took place inside GovTech's managed Google Cloud Platform environment, keeping training data isolated. Under No Citizen Data, training strictly excluded private or personally identifiable information, relying instead on anonymised datasets drawn from scam reports and intelligence feeds. Under Stress-Testing, the model was tested against synthetic dialogues and local corpora covering diverse scam categories and evolving impersonation tactics.
"Fighting scams shouldn't mean sacrificing personal privacy," said Mark Chen, Director of Trust and Safety at GovTech Singapore, in the announcement. "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 completely private."
Part of a Broader Google-Singapore AI Partnership
The scam detection model was delivered as part of Google's National AI Partnership with the Singapore Government, which Google DeepMind announced on May 20, 2026, after opening a research lab in Singapore the previous year. That partnership includes a range of other programs: an exploration of AI co-clinician collaboration with public health clusters, support for infectious disease research through a $7 million Google.org contribution to the Philanthropy Asia Alliance's Health for Human Potential coalition, a Gemma-powered running assistant for blind and low-vision athletes tested with SG Enable, training for local researchers on agentic AI science tools with the National Research Foundation, Gemini for Education for educators, and research on multimodal and multilingual safety benchmarks with the Infocomm Media Development Authority and MLCommons.
An Integration Path for Messaging Platforms
Google DeepMind says messaging platforms and operating system vendors can directly integrate these lightweight models to safeguard users locally, keeping chat interactions secure without exposing private data. The announcement positions the project as a template for messaging platforms balancing user protection with end-to-end encryption — and as an example of governments and model providers working together on safety problems that neither can solve alone.
For a region where scam losses run into the billions annually, an on-device guardrail that respects encryption could prove consequential. Whether other governments follow Singapore's lead will depend on how the model performs once it reaches real users' phones.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
