DeepMind and Singapore launch on-device scam detection model

Google DeepMind and the Government Technology Agency of Singapore (GovTech) announced a joint effort to address scams that account for over half of reported crimes in Singapore and more than S$4 billion in losses since 2019. The partnership produced a fine‑tuned version of DeepMind’s lightweight Gemma model, trained on local scam conversation data to distinguish benign chatter from manipulative intent.

The model runs entirely on the device, processing conversation data only with the user’s consent and without sending messages to cloud servers. It uses a two‑stage pipeline: an efficient classifier screens inbound messages, followed by an in‑context analysis with the fine‑tuned Gemma model to confirm threats. Development took place in GovTech’s managed GCP environment, excluded private data, and relied on anonymised scam reports and synthetic dialogues for testing.

For AI learners, the project illustrates how edge AI can be engineered for real‑time safety while preserving privacy, offering a concrete example of on‑device model deployment and strict data governance.