Google DeepMind and GovTech unveil on-device Gemma scam detection for Singapore
Google DeepMind and the Government Technology Agency of Singapore (GovTech) have announced an on-device scam detection model, built under their National AI Partnership with the Singapore Government. The work was presented at Singapore International Cyber Week (SICW 2026).
The problem is large. Scams make up over half of Singapore's reported crimes and account for more than S$4 billion in losses since 2019. Tactics have moved from simple phishing to complex social engineering, where attackers win a victim's trust through convincing conversations before asking for money. Such scams now account for up to 80% of all losses.
The model starts from DeepMind's lightweight Gemma model, which was fine-tuned on local scam conversations. These ranged from impersonations of public institutions, banks and local figures to trending regional schemes. The aim is a model that can tell benign social chatter from genuine manipulation, something generic models struggle with when deception is localised.
The key design choice is that conversation data is processed on the device. Traditional threat detection relies on server-side scanning of messages in the cloud, but many messaging platforms are end-to-end encrypted, so they cannot inspect private conversations. The announcement's answer is to bring the threat analysis to the device. When the model detects manipulative conversational patterns, it triggers a real-time warning on the user's device, advising them to pause and be cautious before sharing sensitive information or funds. The model is optimized to run on typical smartphones in Singapore.
Because battery efficiency is a critical concern, the system uses a two-stage pipeline. Stage 1 is a lightweight, efficient classifier that screens inbound messages using minimal power and tries to catch potential threats. Stage 2 is a contextual evaluation by the fine-tuned Gemma model, which analyses a larger snapshot of the conversation to confirm a threat or discard a false alarm.
The announcement lists three governance measures. First, a "Sovereign Sandbox": fine-tuning took place inside GovTech's managed GCP environment, keeping training data isolated. Second, "No Citizen Data": training strictly excluded private or personally identifiable information and used anonymised datasets derived from scam reports and intelligence feeds. Third, stress-testing: the model was tested against synthetic dialogues and local corpora covering diverse scam categories and evolving impersonation tactics.
The authors present the project as a model for messaging platforms balancing user protection with end-to-end encryption, and say messaging platforms and operating system vendors can directly integrate these lightweight models to protect users locally.
Key facts
- Google DeepMind and GovTech announced an on-device scam detection model built by fine-tuning the lightweight Gemma model on local Singapore scam conversations.
- Scams make up over half of Singapore's reported crimes and have caused more than S$4 billion in losses since 2019; complex social engineering scams account for up to 80% of losses.
- Conversation data is processed on the device rather than scanned in the cloud, which matters because many messaging platforms are end-to-end encrypted.
- A two-stage pipeline keeps battery use down: a lightweight classifier screens messages, then the fine-tuned Gemma model examines a larger snapshot of the conversation.
- Training used anonymised datasets from scam reports and intelligence feeds, inside GovTech's managed GCP environment, with no private or personally identifiable information.
Why it matters
Scam detection normally depends on scanning messages on a server, which end-to-end encrypted platforms cannot do. This announcement shows another route: run the analysis on the phone, so the conversation stays private and the user still gets a warning. The scale of the problem is the motivation: over half of Singapore's reported crimes are scams, with more than S$4 billion lost since 2019, and up to 80% of losses come from conversational social engineering rather than simple phishing.
Who it affects
The immediate audience is people in Singapore using typical smartphones, the devices the model is optimized for. The source also addresses messaging platforms and operating system vendors, saying they can directly integrate these lightweight models to protect users locally. It frames the project as an example of governments and model providers building technology for public good.
How to use it
There is nothing for an individual to install or switch on yet. No release date, rollout plan or availability to the public is stated; the source does not say which apps or platforms will use the model. For platform and OS teams, the described pattern is a cheap first-stage classifier that gates a heavier fine-tuned model, plus on-device inference so encrypted chats never leave the phone.
How solid is it
This is a first-party announcement from Google DeepMind about its own project with GovTech. The governance steps are described in concrete terms: an isolated GCP environment, anonymised training data, and testing against synthetic dialogues and local corpora. What is missing is evidence of performance: no detection accuracy, false-positive rate or any other performance metric is given.
Risks and caveats
No detection accuracy or false-positive rate is published, so it is unclear how often the model catches real scams or raises false alarms. No battery consumption or latency figures are given for either stage, even though battery efficiency is named as a critical concern. The model issues a warning to the user; the source does not say it blocks messages. The Gemma version and model size are not stated, and no messaging platform, phone maker or operating system vendor is named as an adopter.
“our solution is to bring the threat analysis to the device”
— Google DeepMind blog post