Google DeepMind and the Government Technology Agency of Singapore have built a fine-tuned artificial intelligence model designed to detect scams directly on consumer devices, Google DeepMind announced on October 14, 2026.
Scams make up more than half of Singapore's reported crimes and account for over S$4 billion in losses since 2019, according to Google DeepMind. Attackers have shifted from simple phishing links to complex social engineering, manipulating victims through extended conversations before requesting money. Those conversational schemes now represent up to 80 percent of all scam losses in the country.
Engineers fine-tuned Google’s lightweight Gemma model on local scam patterns, including impersonations of public institutions, banks, and public figures, alongside regional fraud schemes. Unlike traditional detection systems that scan messages on cloud servers, the fine-tuned software processes conversation data locally on the user's device.
On-device screening and governance
The system relies on a two-stage pipeline to limit battery drain on mobile devices. A lightweight initial classifier screens incoming messages using minimal power to flag potential threats. When it detects suspicious activity, the fine-tuned Gemma model performs an in-context analysis on a larger snapshot of the exchange to confirm manipulation or dismiss a false alarm. If confirmed, the software triggers an on-device alert advising the user to pause before sending funds or credentials.
GovTech Singapore ran the fine-tuning process within its own managed Google Cloud Platform environment to keep training data isolated. The project excluded citizen personal data, relying instead on anonymised scam reports, intelligence feeds, and synthetic dialogues. Google DeepMind presented the work at Singapore International Cyber Week 2026, noting that the model requires user consent and operates without transmitting conversation records to either Google or the Singapore government.
