The system uses a two-stage pipeline: a low-power classifier screens inbound messages, then Gemma evaluates a larger conversation snapshot to confirm a threat or discard a false alarm. It issues a real-time warning on the user’s device. DeepMind says training took place in a GovTech-managed GCP environment, excluded private or personally identifiable citizen data, and used anonymised datasets derived from scam reports and intelligence feeds. The source presents this as a way to analyze manipulative conversation patterns without server-side scanning of private chats. It does not report deployment scale, detection accuracy, or false-positive rates.
Raw SignalOct 14, 20261 min read1 source
Google DeepMind and Singapore Government announce new AI tool to protect citizens against scams
In one minute
Google DeepMind and GovTech Singapore have developed an on-device scam-detection model based on a lightweight Gemma model fine-tuned on local scam conversations.