On-Device AI Against Scams: Google DeepMind & MDDI

This post details the fine-tuning of a lightweight Gemma model for on-device scam detection, specifically tailored to local Singaporean scam tactics. It introduces a two-stage pipeline for efficient processing on smartphones: a lightweight classifier for initial screening and a more in-depth analysis with the fine-tuned Gemma model for confirmation. The development process adhered to strict governance, including training within a Sovereign Sandbox, exclusion of citizen data, and stress-testing against synthetic dialogues and local corpora.