Google DeepMind and Singapore GovTech Build On-Device Anti-Scam AI Model with Gemma

GovTech Singapore and Google DeepMind Partnership at SICW 2026

At the Singapore International Cyber Week (SICW) on October 14, 2026, Singapore Minister for Digital Development and Information Josephine Teo announced a strategic partnership between Singapore's Government Technology Agency (GovTech) and Google DeepMind. The joint initiative focuses on developing an on-device artificial intelligence model based on Gemma to detect social engineering scams in real time.

This project addresses an escalating fraud problem in Singapore. Since 2019, scam losses in the country have exceeded 4 billion Singapore dollars. More than 80 percent of these financial losses stem from conversational manipulation and social engineering rather than traditional technical malware or system exploits.

Overcoming End-to-End Encryption Constraints with Local Processing

Most conversational scams operate across messaging applications protected by end-to-end encryption (E2EE), such as WhatsApp and Telegram. While encryption prevents network interception, it also blocks cloud-based security scanners from evaluating message contents for fraudulent behavior.

The model developed by Google DeepMind and GovTech resolves this limitation by running inference directly on local smartphone hardware. The architecture operates under a strict zero cloud data egress policy. Because conversation analysis takes place inside the phone memory, private messages never leave the device or route to external servers. This design preserves user privacy while enabling real-time threat detection.

Two-Stage Processing Pipeline for Power Efficiency

Running language models continuously on mobile hardware drains battery life and strains neural processing units (NPUs). To maintain power efficiency, DeepMind and GovTech implemented a two-stage processing pipeline:

• Lightweight Screening Classifier: A compact, low-power binary classifier continuously monitors incoming notifications and messages. This module flags anomalous triggers while consuming negligible battery power.

• Contextual Gemma Evaluation: When the screening layer detects potential anomalies, the system activates the fine-tuned Gemma model. This model analyzes broader conversational context to identify high-pressure tactics, fabricated urgency, impersonation claims, or demands for immediate funds transfer.

Model Fine-Tuning on Local Threat Intelligence

Google DeepMind fine-tuned the lightweight open Gemma model using anonymized datasets of scam conversations provided by GovTech. The training process occurred within GovTech's sovereign Google Cloud Platform sandbox environment.

The training corpus covers regional fraud schemes, including impersonations of local banks, the Singapore Police Force (SPF), the Inland Revenue Authority of Singapore (IRAS), and fraudulent investment schemes. GovTech stripped all personally identifiable information (PII) from the data before the training pipeline ingested it.

Integration with the National ScamShield Ecosystem

The on-device model will connect with Singapore's national anti-scam infrastructure, particularly the ScamShield application managed by GovTech and law enforcement agencies.

When the system detects manipulative interaction patterns, it generates an on-screen warning directly on the device. This real-time intervention prompts users to pause and verify requests before transferring funds or disclosing authentication credentials.