StorySoftware
Google DeepMind and Singapore's GovTech build an on-device model to flag scam conversations
Google DeepMind and Singapore's GovTech are fine-tuning Google's lightweight Gemma model to spot scams in messages. The model runs on the phone and warns users to pause before sharing information or money. It is not yet broadly available, and rollout inside apps will be phased.
- Trained on local scam conversations, including impersonation of public institutions, banks and local figures.
- Two stages: a lightweight classifier screens messages cheaply, then the Gemma model analyses a larger snapshot to confirm or dismiss a threat.
- Investment fraud, fake job offers and official impersonation make up over half of Singapore's reported crimes; losses exceed S$4 billion since 2019.
- No test results are published; stress tests used synthetic dialogues and local corpora.
- Conversations are accessed only with user consent and under each app's terms, according to the post.