The conversation is judged on the phone
Google DeepMind and the Government Technology Agency of Singapore, GovTech, described an on-device scam detection model. They say it is optimized to run on typical smartphones in Singapore. The model processes conversation data on the device. Many chats are end-to-end encrypted, so a cloud scanner cannot read them, and the check stays on the phone.
Scams make up over half of Singapore's reported crimes and account for more than S$4 billion in losses since 2019. The tactics have moved from simple phishing to social engineering, a convincing conversation that ends in a request for money. Those scams now account for up to 80% of all losses. They say generic models often miss local deception, so a lightweight Gemma model was fine-tuned on local scam conversations, including impersonation of public institutions, banks, and local figures, and regional schemes.
[1]
Screen one message, then read a longer stretch
The post describes two stages. A lightweight classifier screens inbound messages, using little power, and looks for possible threats. A fine-tuned Gemma model then reads a larger snapshot of the conversation, confirms a threat, or discards a false alarm. When it sees manipulative patterns, it sends a real-time warning on the device, telling the person to pause before sharing sensitive information or funds.
The collaboration was presented at Singapore International Cyber Week 2026. Fine-tuning took place in a GCP environment managed and maintained by GovTech, which they call a Sovereign Sandbox, with the training data kept isolated.
[1]Training excluded citizen data. Accuracy is not in the post
Training strictly excluded private or personally identifiable information. The sets were anonymised material from scam reports and intelligence feeds. Tests used synthetic dialogues and local corpora, covering different scam categories and changing impersonation tactics.
The post gives no accuracy, no false-alarm rate, and no date or phone model for installation. A quote from Mark Chen, Director of Trust and Safety at GovTech Singapore, restates the same point: catch local scams on the device and keep personal communications private. It is not an evaluation table.
[1]