Google DeepMind and the Government Technology Agency of Singapore (GovTech) have jointly developed an on-device scam detection model tuned to local deception tactics and optimized to run on typical smartphones. The work was carried out under the National AI Partnership with the Singapore Government and was presented at Singapore International Cyber Week (SICW 2026).
The problem being addressed
Scams represent more than half of reported crimes in Singapore and have caused over S$4 billion in losses since 2019. Attackers have shifted from simple phishing to sophisticated social engineering, where convincing conversations are used to manipulate victims before requesting money. These social engineering scams now account for up to 80% of total losses, creating an urgent need for localized and privacy-preserving protections.
How the solution works
The team fine-tuned Google DeepMind’s lightweight Gemma model with locally relevant scam conversations, including impersonations of public institutions, banks and local figures as well as regionally trending schemes. The intent was to help the model distinguish innocent social chatter from manipulative tactics.
Unlike traditional server-side threat detection, this system performs analysis on the device. Many messaging platforms use end-to-end encryption, preventing cloud-side inspection of private conversations. Bringing threat analysis to the device enables faster detection of manipulative conversational patterns and allows the user to receive a warning before financial or credential exposure.
Technical design and efficiency
Running a capable scam classifier on modern smartphones requires energy-efficient design. The deployed solution uses a two-stage pipeline:
- Stage 1: A lightweight, efficient classifier screens incoming messages with minimal power consumption to flag potential threats.
- Stage 2: When the first stage finds a signal, the fine-tuned Gemma model evaluates a larger conversation snapshot to confirm a threat or rule out a false positive.
If manipulative patterns are detected, the model issues a real-time warning on the user’s device advising caution before sharing sensitive information or funds.
Governance and data handling
Development was governed by strict controls to preserve sovereignty and privacy:
- Sovereign Sandbox: Fine-tuning took place within GovTech’s managed Google Cloud Platform environment to keep training processes isolated.
- No Citizen Data: Training excluded private or personally identifiable information; datasets were anonymized and derived from scam reports and intelligence feeds.
- Stress-Testing: The model was tested against synthetic dialogues and local corpora to cover diverse scam categories and evolving impersonation tactics.
Implications and next steps
The project demonstrates that on-device AI can help messaging services counter criminal scams without compromising end-to-end encryption or exposing private data. It also illustrates a collaboration model between a government agency and a model provider to produce technology intended for public benefit. Messaging platforms and operating system vendors can integrate these lightweight models locally to help safeguard users and reduce fraud while keeping chats private.
