DeepMind and Singapore Put Scam Detection on Phones, Not in the Cloud
Google DeepMind and GovTech built an on-device scam detector using Gemma that catches social engineering without uploading your chats.
Models · Source: Google DeepMind
What happened
Google DeepMind partnered with GovTech Singapore to build an on-device scam detection model using a fine-tuned version of Google's lightweight Gemma model. The system analyzes conversational patterns in real time to spot manipulative social engineering before users lose money or surrender credentials. Crucially, the model runs entirely on the smartphone, ensuring private messages are never transmitted to cloud servers.
The initiative tackles a massive national problem. Scams account for more than half of Singapore's reported crimes and have drained over S$4 billion since 2019. Modern attackers have shifted from crude phishing links to long, convincing social engineering conversations, impersonating banks and public officials. These multi-step dialogues now drive up to 80 percent of all scam losses in the country, making traditional keyword blockers and generic cloud models ineffective.
Running language models continuously on mobile hardware drains batteries quickly. To solve this, the engineering teams built a two-stage pipeline. A low-power, lightweight classifier screens incoming inbound text first. If it flags suspicious conversational dynamics, the fine-tuned Gemma model evaluates a wider contextual snapshot to confirm or discard the threat. The model operates solely with user consent, trained inside GovTech's sovereign cloud sandbox using synthetic dialogues and anonymized threat reports without touching citizen data.
Key facts
- S$4 billion — Scam losses in Singapore since 2019
- 80% — Share of scam losses driven by social engineering conversations
- 2-stage — Pipeline screening messages to preserve mobile battery
- Gemma — Lightweight Google model fine-tuned for local on-device inference
Why it matters
This project proves you can deliver high-context AI safety inside end-to-end encrypted environments. Cloud-hosted moderation APIs are fundamentally incompatible with private messaging architectures because scanning requires breaking encryption or shipping raw text to third-party servers. Pushing specialized, quantized models to the edge preserves client privacy while maintaining real-time safety interventions.
It also exposes the limits of generic frontier models in defense applications. Global models fail against localized scams that exploit regional slang, local banks, and country-specific agency names. Fine-tuning smaller open weights like Gemma on isolated, localized threat intelligence delivers superior accuracy with zero server inference bills and zero cloud data exposure.
For builders
Use two-stage pipelines for edge inference
Never run a full model on every incoming background event. Pair a minimal heuristic or linear classifier with a quantized SLM like Gemma that only wakes up when confidence drops below safe thresholds. This protects device battery life and avoids thermal throttling.
Protect encrypted workflows without cloud leaks
If you build consumer communication tools, enterprise compliance will reject routing private chat history to remote APIs. Delivering fine-tuned on-device models allows you to offer proactive fraud and compliance guardrails without violating end-to-end encryption guarantees.
Local fine-tuning beats massive generic models
Targeted fraud uses hyper-local social engineering that broad foundational models routinely miss. Fine-tuning small open-weight models on proprietary or localized threat corpora yields higher precision at virtually no recurring marginal cost.
My take
Everyone obsessed with massive frontier models in the cloud is missing where consumer trust will actually be won. The most effective security AI will never watch you from a remote data center. It will sit directly on your silicon, guard your conversations locally, and report back to absolutely nobody.
Original reporting: Google DeepMind. This is my rewrite and opinion.