- Published on
Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection
- Authors
- Name
- DeepHorus
Google DeepMind and GovTech Singapore have fine-tuned a privacy-preserving, on-device AI model to detect scams. Part of the National AI Partnership, it was presented at Singapore International Cyber Week (SICW 2026). It protects users without giving governments or Google access to private conversations.
Scams account for over half of Singapore’s reported crimes and more than S$4 billion in losses since 2019. Tactics have shifted from phishing to complex social engineering, now up to 80% of all losses. Attackers impersonate officials, banks, and local figures to manipulate victims.
The model is built on Google’s lightweight Gemma and fine-tuned on local scam conversations. It can distinguish benign social chatter from genuine manipulation when deployed in citizen applications. Unlike cloud-based scanning, it processes conversation data directly on-device, enabling faster detection of manipulative intent.
A two-stage pipeline manages battery efficiency. Stage one uses a lightweight classifier to screen inbound messages with minimal power. Stage two analyzes a larger conversation snapshot with Gemma to confirm threats or discard false alarms. A real-time warning advises users to pause before sharing sensitive information or funds.
Governance remained strict. Fine-tuning occurred in GovTech Singapore’s sovereign sandbox on Google Cloud, keeping training data isolated. No private or personally identifiable information was used; datasets were anonymized from scam reports and intelligence feeds. The model was stress-tested against synthetic dialogues and local corpora.
Deployment will be carefully phased with the Singapore government and only with user consent. Because the model runs entirely on-device, neither the government nor Google can access users’ conversations. The project offers a model for messaging platforms balancing protection with end-to-end encryption.
The collaboration shows how AI can counter criminality while preserving privacy and public trust. It also demonstrates how governments and model providers can build technology for public good. Further rollout details will emerge as applications integrate the model.
