Google DeepMind’s on-device scam detection keeps chats off the cloud

What it claims

Google DeepMind and GovTech Singapore built an on-device scam detector for messaging apps. It is pitched as a way to protect users without breaking end-to-end encryption, because no chat content leaves the phone. The target is conversational social-engineering scams, not one-off phishing links. The page says these make up up to 80% of Singapore's scam losses. It also says scams are more than half of the country's reported crimes, with over S$4 billion lost since 2019. The work was announced in an October 14, 2026 DeepMind blog post and presented at Singapore International Cyber Week (SICW 2026). This page is a secondary, AI-assisted news write-up of that post, not the post itself.

How it works

flowchart LR
  subgraph Device["User device (consented chats only)"]
    M[Incoming message] --> C[Low-power classifier]
    C -->|not suspicious| X[No action]
    C -->|flagged| G[Fine-tuned Gemma reads conversation]
    G -->|false alarm| X
    G -->|confirmed| W[Real-time warning]
  end
  subgraph GCP["GovTech-managed GCP"]
    D[Anonymised scam reports + intel feeds] --> F[Gemma fine-tuning]
  end
  F -.model.-> G

The design choice that matters is the gate. The cheap classifier runs on every message, and Gemma only reads the wider conversation when the classifier flags something. The page presents this as the fix for the battery cost of running a full scam model on a phone. Gemma's job is to cut false positives, not to catch scams the classifier missed.

Testing used two sources:

• synthetic dialogues
• local corpora covering different scam categories and evolving impersonation tactics

Limits

• Nothing you can adopt yet. No weights, licence, API or SDK are mentioned. Rollout inside applications will be "phased" and developed with the Singapore Government.
• No model sizes. The page names neither the classifier nor the Gemma variant or parameter count.
• No evaluation results. The test sets are named, but there are no precision, recall or false-positive rates, no latency or battery figures, and no comparison against a baseline. Whether the cascade actually saves power, or a localised model beats a generic one, is asserted rather than shown.
• Singapore-specific. The fine-tuning targets local institutions, banks, public figures and regional schemes. Nothing is said about other markets or languages.
• Consent dependency. The model sees chats only with user consent, under each host app's terms of use. Coverage therefore depends on opt-in.
• For anything beyond this, go to the original DeepMind blog post. This article adds no technical detail of its own.