Google DeepMind, GovTech Singapore Unveil Private On-Device Scam Detector
Google DeepMind and GovTech Singapore unveil private, on-device scam detection as fraud exceeds S$4 billion and causes over half of all reported crimes.
Summary
Google DeepMind and Singapore’s Government Technology Agency, GovTech, announced an on-device scam detector on October 14, 2026, through their National AI Partnership. Scams exceed half of Singapore’s reported crimes and have caused more than S$4 billion in losses since 2019. Social engineering, in which attackers build convincing conversations before requesting money, now causes up to 80% of losses. The lightweight Gemma model was fine-tuned to distinguish benign chat from regional schemes and impersonations of public institutions, banks and local figures.
Optimized for typical smartphones in Singapore, the model analyzes conversations locally because end-to-end encryption prevents messaging platforms from scanning private chats in the cloud. A low-power classifier first screens inbound messages, then Gemma examines a larger conversational snapshot to confirm threats or discard false alarms. Detected manipulation triggers a real-time warning advising users to pause before sharing funds, credentials or sensitive information. GovTech trust and safety director Mark Chen said this provides targeted local protection while keeping communications private.
Fine-tuning occurred inside GovTech’s managed GCP sovereign sandbox, isolating training data. Private and personally identifiable citizen information was excluded, with anonymized scam reports and intelligence feeds used instead. Testing covered synthetic dialogues, local corpora, diverse scam categories and evolving impersonation tactics. Presented at Singapore International Cyber Week 2026, the model can be integrated by messaging platforms and operating system vendors, although no public deployment date or platform integration was specified.
Positives
- On-device analysis protects end-to-end encrypted conversations without sending private messages to cloud scanners.
- Gemma was fine-tuned for Singaporean schemes, including impersonations of public institutions, banks and local figures.
- A two-stage pipeline combines low-power screening with contextual verification to limit battery use and reject false alarms.
- Training excluded private and personally identifiable citizen data, relying on anonymized scam reports and intelligence feeds.
- Messaging platforms and operating system vendors can integrate the lightweight model for local user protection.
Risks & concerns
- Scams account for over half of Singapore’s reported crimes and more than S$4 billion in losses since 2019.
- Social engineering scams now generate up to 80% of losses by manipulating victims through convincing conversations.
- Generic models struggle with localized deception techniques, requiring country-specific fine-tuning and testing.
- Battery efficiency remains a critical constraint when running complex scam classification on smartphones.
- No public deployment date or confirmed messaging platform or operating system integration was specified.