DeepMind and Singapore Train Model to Catch Scammers

🛡 DeepMind and Singapore Train Model to Catch Scammers

At Singapore International Cyber Week 2026, Google DeepMind and GovTech Singapore presented a scam communication detector that runs on the user's device. It is based on the open-source Gemma model, fine-tuned on anonymized reports.

🌍 First public example: a government fine-tuned an open LLM in its own 'sovereign' sandbox and deployed it on citizens' devices without vendor access to the communications. A cascade with a lightweight classifier saves battery: Gemma activates only upon suspicion.

👤 Messages do not leave the phone, and the model reads them only with your consent. In case of an alert, it warns not to rush with money and data. Specific apps and timelines were not named; the launch will be phased.

Source 1: https://deepmind.google/blog/on-device-ai-scam-detection-mddi/ Source 2: https://www.neoteo.com/en/google-deepmind-and-govtech-singapore-plan-on-device-gemma-scam-detection