On-device scam detection: DeepMind and Singapore fine-tune Gemma
Google DeepMind and GovTech Singapore have fine-tuned a lightweight Gemma model that flags conversational scams directly on the phone — without chat content leaving the device. What is documented, how the two-stage pipeline works, and what enterprises can take away.
CogLake Editorial · 14 October 2026 · 4 min
Transparency: This article was automatically generated with the assistance of AI. All statements are based on the primary sources linked below, retrieved on October 14, 2026. Figures on scam losses come from the sources cited by Google.
On-device scam detection: DeepMind and Singapore fine-tune Gemma
In one sentence: On October 14, 2026, Google DeepMind and GovTech Singapore announced a scam detection model, fine-tuned from Google's lightweight Gemma family, that analyses conversations directly on the smartphone — so chat content does not leave the device.
What was announced
The two organisations presented the project at Singapore International Cyber Week (SICW 2026). It builds on the national AI partnership between Google and Singapore. The motivation, per the announcement: scams make up more than half of reported crimes in Singapore and more than S$4 billion in losses since 2019, and conversational social engineering now accounts for up to 80 percent of losses (as cited from a Straits Times tracker).
Unlike classic threat detection, which scans messages server-side in the cloud, this model is to run inside apps installed by citizens. According to Google, neither the Singapore government nor Google can read users' conversations through its deployment; access to on-device conversations requires the user's consent and follows each app's terms of use.
The two-stage pipeline
| Stage | Task | Stated purpose |
|---|---|---|
| 1. Lightweight classifier | Screens inbound messages with minimal power | Catches potential threats cheaply and protects battery life |
| 2. Fine-tuned Gemma model | Analyses a larger snapshot of the conversation in context | Confirms the threat or discards a false alarm and triggers an on-device warning |
The Gemma model was fine-tuned on local scam conversations: impersonation of public institutions, banks and local figures, as well as regionally typical schemes. That local context is the point — generic models, the authors argue, often miss local deception techniques.
What is documented — and what remains open
| Statement | Status |
|---|---|
| DeepMind × GovTech collaboration, on-device deployment, consent requirement | Documented in the announcement |
| Two-stage pipeline (classifier + Gemma), real-time warning on the device | Documented in the announcement |
| Governance: training in a GovTech-managed GCP environment ("Sovereign Sandbox"), no citizen data, anonymised training sets, stress tests with synthetic dialogues | Documented in the announcement |
| Which Gemma version and size exactly | Not stated |
| Accuracy, false-positive rates, rollout date and target apps | Not stated; rollout is to be "carefully phased" |
Why it matters for enterprises
Three takeaways transfer beyond Singapore:
- Local processing is a privacy pattern, not a niche. Wherever chat or message content is too sensitive for cloud models, on-device or on-premise processing keeps data under your own control.
- A small model plus a classifier can be enough. The two-stage design shows a practical cost pattern: a cheap filter decides when the expensive model runs.
- Governance is part of the engineering. Isolated training environment, no personal data in training, stress tests — the announcement treats these as first-class requirements.
For CogLake customers the pattern is familiar: local models are available in correspondingly configured dedicated and on-premise deployments, and the privacy layer at the AI boundary can mask or block sensitive data before it reaches any model — regardless of where that model runs. See CogLake Security and AI models.
Sources
- Google DeepMind blog: "Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection", October 14, 2026 — deepmind.google/blog/on-device-ai-scam-detection-mddi/ (retrieved October 14, 2026)
- Straits Times scam tracker (loss figures, as cited by Google) — retrieved October 14, 2026
- Background: "Strengthening Singapore's AI Future: A New National Partnership", May 2026 — deepmind.google/blog/strengthening-singapores-ai-future-a-new-national-partnership/