Google DeepMind and Singapore Government build on-device scam detection model using fine-tuned Gemma

Google DeepMind and the Government Technology Agency of Singapore (GovTech Singapore) collaborated to fine-tune a lightweight Gemma model for on-device scam detection, aiming to intercept conversational scams in private chats without sending data to the cloud.

Scams now account for up to 80% of financial losses in Singapore, with over half of reported crimes involving scams and more than S$4 billion in losses since 2019.

The fine-tuned model was trained on anonymized datasets derived from scam reports and intelligence feeds, excluding private or personally identifiable information, and was developed within GovTech Singapore’s managed GCP environment.

The model processes conversations entirely on-device, triggering real-time warnings when manipulative patterns are detected, and will only access user conversations with explicit consent under the terms of the hosting application.

The deployment uses a two-stage pipeline: a lightweight classifier screens inbound messages for potential threats with minimal power usage, followed by in-context analysis with the fine-tuned Gemma model to confirm or discard threats.

The collaboration was presented at Singapore International Cyber Week 2026 and emphasizes strict governance, including sovereign sandboxing of training processes and exclusion of citizen data from training datasets.