DeepMind and Singapore Test On-Device Scam Detection

A fine-tuned Gemma model spots manipulation in private chat messages directly on hardware, bypassing the need to route private text to cloud servers.

Google DeepMind and GovTech, Singapore’s government technology agency, unveiled a joint initiative on Wednesday (October 14, 2026) testing an on-device scam detection system designed to catch fraud attempts directly on mobile phones. Presented at the Singapore International Cyber Week (SICW 2026), the project adapts Google’s open Gemma model family to analyze chat context locally, preventing message logs from ever touching external servers.

Scams have become a primary public security headache in Singapore. Official records show fraudulent schemes now account for over 50% of all reported crimes in the country. Since 2019, cumulative losses have crossed 4 billion Singapore dollars (roughly 3 billion US dollars). Social engineering, where attackers trick victims into authorizing money transfers themselves, causes up to 80% of that financial damage.

The encryption dilemma

Stopping modern social engineering usually breaks against modern privacy standards. Mainstream chat apps like WhatsApp and Signal rely on end-to-end encryption. Only the sender and the receiver hold the decryption keys. Because platform operators cannot read message contents on central servers, cloud-based spam filters cannot inspect suspicious conversations.

DeepMind and Singaporean engineers bypassed this barrier by shifting analysis to the handset. Instead of asking platforms to inspect encrypted traffic in transit, the model evaluates text locally as it renders on the user’s screen.

How on-device scam detection works

Running generative models on phones drains battery life and creates thermal bottlenecks. To keep resource use low, the system splits work across two stages.

In the first stage, a lightweight classifier runs in the background. It consumes negligible power while checking incoming messages for suspicious wording. If this initial filter spots common scam indicators, it triggers the second stage: a specialized Gemma model tuned to evaluate conversational context. This larger model examines whether psychological manipulation or coercion is occurring and pushes an instant alert to the screen. For regular day-to-day texts, the secondary model stays dormant.

Sovereign training and synthetic data

The tool requires explicit user consent and runs entirely on device silicon. Neither Google nor GovTech collects message transcripts, generated alerts, or device identities.

Training the system on regional fraud tactics without compromising citizen privacy required an isolated setup. GovTech and DeepMind ran model training inside a sovereign sandbox environment on Google Cloud. The dataset excluded personal citizen communications. Instead, engineers trained the weights on anonymized police incident reports, threat intelligence feeds, and synthetic scam dialogues generated to simulate realistic local fraud pitches. Mark Chen, director of Trust & Safety at GovTech, stated that defending against regional scams should not require citizens to give up communication privacy.

What remains unannounced

While the demonstration confirmed the architecture at SICW 2026, DeepMind and GovTech have not shared formal benchmark results, such as latency numbers or false-positive rates.

Neither team has announced a public release timeline. How the software will reach everyday users also remains unresolved: authorities have not confirmed whether it will ship inside Singapore’s ScamShield app, integrate directly into Android, or exist as an optional software library for third-party messaging clients.