Google DeepMind Builds On-Device Scam Detection With Singapore

Google DeepMind Singapore Scam Detection Partnership

Google DeepMind announced on October 14, 2026, that it has partnered with the Government Technology Agency of Singapore to create a fine-tuned artificial intelligence model designed to spot conversational scams directly on users' devices. The collaboration falls under a broader National AI Partnership between the research lab and the Singapore government. According to the announcement, the system uses a lightweight version of Google's open Gemma model that has been adapted to recognize local fraud patterns without ever sending message data to the cloud.

The move responds to a sharp rise in socially engineered scams across the city-state. Official figures show that scams now account for more than half of all reported crimes in Singapore and have drained upwards of S$4 billion from victims since 2019. The announcement notes that tactics have shifted from simple phishing links to extended conversations in which attackers build trust before asking for money or credentials. Those complex schemes now drive roughly 80 percent of total financial losses.

On-Device vs Server-Side Scam Detection

Traditional server-side scanning would require copying private chats to a central system, a trade-off many users and platforms reject. The new approach keeps every analysis step on the phone itself. When the model flags a manipulative pattern, it pushes a real-time warning to the user suggesting they pause before sharing sensitive information or transferring funds. Because the inference happens locally, neither Google nor the Singapore government can view the underlying conversations through this deployment.

Engineers fine-tuned the Gemma variant on anonymized datasets drawn from scam reports and intelligence feeds rather than personal messages. The training corpus covers impersonations of public agencies, banks, and local figures as well as trending regional schemes. That local context helps the model separate ordinary chatter from coordinated manipulation attempts that generic classifiers often miss. The announcement says the rollout inside citizen-facing applications will be phased and coordinated with GovTech, and the model will only access on-device conversations after the user grants explicit consent in line with each app's terms of use.

Two-Stage Pipeline for Mobile AI

Running a capable classifier on a smartphone battery budget required a two-stage pipeline. A first, ultra-lightweight filter scans incoming messages using minimal power to surface potential threats. Only when that filter trips does the fine-tuned Gemma model engage, examining a wider slice of the dialogue to confirm or dismiss the alert. The design aims to keep false positives low while preserving battery life for everyday use.

Governance controls were baked into the development process. Fine-tuning took place inside a sovereign sandbox, a GovTech-managed Google Cloud environment that isolates training data. No personally identifiable information entered the pipeline, and the model was stress-tested against synthetic dialogues and local corpora covering diverse scam categories and evolving impersonation tactics. The work was presented at Singapore International Cyber Week 2026 as a reference architecture for messaging platforms that want to add safety tooling without breaking end-to-end encryption.

Privacy-Preserving AI for Messaging Apps

Mark Chen, director of trust and safety at GovTech Singapore, said the project shows that advanced AI protection can reach citizens' devices without sacrificing personal privacy. The announcement frames the effort as a template for how governments and model providers can cooperate on public-good technology that respects encryption and data sovereignty.

For developers building messaging or transaction apps, the case study demonstrates that on-device inference pipelines can now handle nuanced natural-language threats at acceptable latency and power cost. Teams evaluating similar features should audit whether their threat models rely on cloud telemetry that conflicts with encryption promises, and they should consider a staged classifier architecture to balance accuracy with resource constraints. Users in Singapore can expect the first integrations to appear in government-backed applications over the coming months, with opt-in controls governing activation.

Source: Google DeepMind