Singapore and Google DeepMind put scam detection on the phone itself

Google DeepMind and GovTech Singapore have built an on-device scam detection model by fine-tuning Google’s lightweight Gemma on local scam conversations, according to an official DeepMind post. The model runs on the phone, so neither party can read users’ chats.

What is new

The work was carried out under Singapore’s National AI Partnership and presented at Singapore International Cyber Week 2026. DeepMind says the model was fine-tuned on local scam conversations, including impersonation of public institutions, banks and local figures. That local focus is the main difference from a generic spam filter, since scammers in Singapore imitate specific agencies and public figures.

The system uses a two-stage pipeline. In the first stage, a lightweight classifier screens incoming messages while using minimal power. Messages it judges safe pass through. Suspicious ones go to the second stage, where the fine-tuned Gemma model examines a larger snapshot of the conversation and either confirms or dismisses the threat. When it detects manipulative patterns, the user receives a real-time warning on the device, prompting them to pause before sharing sensitive information or sending money.

Privacy is central to the design. The model runs on-device and, according to the post, accesses conversations only with the user’s consent. Neither the Singapore Government nor Google can see users’ conversations through it. Mark Chen, Director of Trust and Safety at GovTech Singapore, said: “Fighting scams shouldn’t mean sacrificing personal privacy.”

On training and governance, DeepMind says fine-tuning took place in GovTech Singapore’s managed Google Cloud environment, with the training data isolated. Private or personally identifiable information was excluded, and the team used anonymized scam reports and intelligence data. The model was stress-tested against synthetic dialogues and local corpora.

The scale of the problem

DeepMind frames the project around Singapore’s scam problem. It says scams such as investment fraud, fake job offers and impersonation of public officials make up more than half of the country’s reported crimes, and have caused losses of more than S$4 billion since 2019. Social engineering, where victims are manipulated into acting, accounts for up to 80% of those losses, according to the post.

Why it matters

The clearest test of this project is whether it moves Singapore’s loss figures, and the baseline is now public. Singapore Police Force data, as reported by Malay Mail, put 2025 scam losses at S$913.1 million, down about 19% from S$1.11 billion in 2024. Investment scams and impersonation of government officials ranked among the top categories by amount lost. Those are the patterns the DeepMind post says the model was tuned for, so a further decline in those categories would be the first outside signal. Any fall will be hard to attribute to one tool, because the model has no stated launch date.

The second thing to watch is the regulatory calendar. Singapore has set new Online Criminal Harms Act codes for seven messaging and conferencing services, including WhatsApp, Telegram and iMessage. According to The Star, compliance is due by 31 January 2027, with measures against impersonation of the Singapore government due earlier, on 30 September 2026. The codes put the duty on platforms, not on a phone-side model. How an on-device warning would sit alongside those platform obligations is not addressed in the post, and neither is whether the model reads chats inside those apps at all.

Three disclosures would confirm or undercut the privacy claim and the usefulness claim: independent measurements of false positives and missed scams, the supported apps and devices, and an account of how consent is requested and revoked. Without them, the promise that no one can read the chats rests on the design description alone.