Google DeepMind and the Government Technology Agency of Singapore have developed an on-device artificial intelligence model to protect smartphone users against scams. The initiative, announced by Google DeepMind, operates as part of the National AI Partnership with the Singapore Government.
Scams make up more than half of all reported crimes in Singapore and account for over S$4 billion in losses since 2019. Fraud tactics have shifted from simple phishing toward social engineering, in which attackers manipulate victims through conversations before requesting money. Google DeepMind reported that these conversational scams now drive up to 80 percent of total scam losses.
Traditional security defenses rely on cloud servers to scan messages, but end-to-end encrypted platforms prevent remote inspection. To protect users without exposing private communications, the new system runs directly on typical smartphones. When the software detects manipulative patterns, it displays a real-time warning on the device advising the user to pause before sharing sensitive credentials or funds.
Detection pipeline
Engineers fine-tuned Google DeepMind's lightweight Gemma model on local scam conversations. The training material covered impersonations of public institutions, banks, and local figures, as well as regional fraud schemes. To protect device battery life, the software uses a two-stage screening pipeline:
- A lightweight classifier screens incoming messages with minimal power draw to flag potential threats.
- The fine-tuned Gemma model evaluates a larger snapshot of the conversation to confirm the danger or dismiss a false alarm.
Governance and testing
GovTech and Google DeepMind presented the tool at Singapore International Cyber Week 2026. Fine-tuning occurred inside GovTech's managed Google Cloud Platform environment to keep training data isolated. The teams excluded citizen records and personally identifiable information from training, relying instead on anonymized scam reports, intelligence feeds, and synthetic dialogues.
Operating system vendors and messaging platforms can directly integrate the lightweight models to screen chat interactions locally without sending private data to external servers.
