Google DeepMind and Singapore GovTech Launch On-Device Gemma Scam Detection Model
Google DeepMind and the Government Technology Agency of Singapore (GovTech) have announced an on-device scam detection model that runs on typical smartphones in Singapore. The tool flags manipulative conversations locally, so it can warn users in real time without scanning private messages on a server — a design that works even when messaging platforms use end-to-end encryption.
The announcement was presented at Singapore International Cyber Week (SICW 2026) and published on the Google DeepMind blog on October 14, 2026, under the lab's Responsibility & Safety category. The post is credited to JJ Geewax and Miguel de Andrés.
The problem: scams are now Singapore's dominant crime category
Scams make up over half of Singapore's reported crimes and account for more than S$4 billion in losses since 2019, according to the announcement. The threat has also changed shape: scam tactics have shifted from simple phishing to complex social engineering, in which attackers manipulate victims through convincing conversations before requesting money. These conversation-based scams now account for up to 80% of all losses.
That shift is what makes detection hard. A classifier tuned to spot a phishing link is not equipped to judge whether a multi-turn conversation is gradually steering a victim toward a transfer.
Why the model runs on the device rather than in the cloud
Traditional threat detection relies on server-side scanning of messages in the cloud. That approach breaks down for end-to-end encrypted messaging platforms, which by design cannot inspect private conversations. Google DeepMind's answer is to move the threat analysis onto the phone itself.
Running locally has two stated benefits: faster analysis of conversation patterns to detect manipulative intent, and the ability for users to act before financial or credential exposure occurs. Because conversation data is processed on the device, the privacy of the chat is preserved.
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 completely private.
— Mark Chen, Director, Trust and Safety, GovTech Singapore
Local context is the reason a generic model is not enough
Generic models often struggle with localised deception techniques, which makes scam detection harder. To close that gap, Google DeepMind fine-tuned its lightweight Gemma model on local scam conversations. The training material covered impersonations of public institutions, banks, and local figures, as well as trending regional schemes. The goal was to let the model distinguish benign social chatter from genuine manipulation.
When the model detects manipulative conversational patterns, it triggers a real-time warning directly on the user's device, advising them to pause and exercise caution before sharing sensitive information or funds.
Architecture: a two-stage pipeline built for battery efficiency
Running a complex scam classifier on a modern smartphone is difficult, and battery efficiency is a critical concern. Google DeepMind addresses this with a two-stage pipeline:
- Stage 1 — lightweight fast screening classifier. A lightweight, efficient classifier screens inbound messages using minimal power, focused on catching potential threats. Messages judged safe receive a benign verdict and stop there.
- Stage 2 — fine-tuned Gemma model. Suspicious messages proceed to contextual evaluation, where the fine-tuned Gemma model analyzes a larger snapshot of the conversation to confirm a threat or discard a false alarm. The output is either a scam alert or a benign verdict.
The two-stage design means the heavier multi-turn analysis only runs on messages that survive the first filter.
Governance constraints on training and testing
Google DeepMind states that strict governance was maintained throughout development, with three specific controls:
- Sovereign sandbox. Fine-tuning took place within GovTech's managed and maintained Google Cloud Platform environment, keeping training data isolated.
- No citizen data. Training strictly excluded private or personally identifiable information, using anonymised datasets derived from scam reports and intelligence feeds.
- Stress-testing. The model was tested against synthetic dialogues and local corpora to cover diverse scam categories and evolving impersonation tactics.
What the collaboration signals for messaging platforms and OS vendors
The project demonstrates that AI can help messaging services counter criminality, and that safety and privacy need not be traded off against each other when protective AI runs directly on-device. It also serves as a template for how governments and model providers can work together to build technology for public good.
Google DeepMind states that 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 Singapore work sits inside the lab's broader National AI Partnership with the Singapore Government, which the blog describes as the arrangement that made the collaboration possible.