Gemma vs. scammers: Google DeepMind and Singapore put fraud detection on the phone itself

Google DeepMind and the Government Technology Agency of Singapore (GovTech) have announced a joint tool meant to shield citizens from scammers without sacrificing the privacy of their messages. Built under the National AI Partnership between DeepMind and the Singapore government, it was presented at Singapore International Cyber Week 2026.

The scale of the problem for the city-state is considerable: scams account for more than half of reported crimes, and losses have topped S$4 billion since 2019. The tactics have evolved, too — from plain phishing to elaborate social engineering, where attackers build a convincing conversation before ever asking for money. Such schemes now drive up to 80% of all losses.

On-device analysis

The core idea is to move threat detection onto the phone itself. Traditional systems scan messages on a server in the cloud, but many messaging apps use end-to-end encryption and simply cannot read private chats. DeepMind's answer sidesteps the conflict: a lightweight Gemma model analyses the conversation locally, so the data never leaves the device.

Generic models handle local deception techniques poorly, so Gemma was fine-tuned on local scam dialogues — from impersonations of public institutions, banks, and well-known local figures to trending regional schemes. That helps it tell harmless social chatter apart from genuine manipulation.

Under the hood

To keep a heavy classifier from draining the battery, the system uses a two-stage pipeline. First, a lightweight, power-frugal filter screens incoming messages. Second, the fine-tuned Gemma model evaluates a larger slice of the conversation to confirm a threat or dismiss a false alarm. When it spots signs of manipulation, the model raises a warning right on the screen, urging the user to pause and be cautious before sharing data or funds.

Built on strict governance

The developers stress their adherence to strict rules. Fine-tuning took place inside a GovTech-managed Google Cloud environment — a "sovereign sandbox" that keeps training data isolated. No private or personally identifiable citizen data was used, only anonymised datasets drawn from scam reports and intelligence feeds. The model was also stress-tested against synthetic dialogues and local corpora to cover diverse scam categories and shifting impersonation tactics.

The authors see the project as a template for messaging services that must balance user protection with end-to-end encryption. Lightweight models like these could be embedded by messaging platforms and operating-system vendors alike, guarding users locally without ever exposing their private data.

Source: Google DeepMind