Google DeepMind and Singapore Government announce new AI tool to protect citizens against scams
Learn how Google DeepMind and GovTech Singapore uses lightweight on-device Gemma models to intercept conversational scams in private chats with zero cloud data egress.
Google DeepMind and the Singapore Government have announced a new AI tool designed to protect citizens from scams by using lightweight on-device models to detect fraudulent conversations without compromising privacy.
What happened
Google DeepMind and the Singapore Government have developed an on-device AI model to detect scam conversations in private chats. This model, based on the Gemma model, is fine-tuned on local scam data and runs directly on smartphones, ensuring privacy and real-time threat detection.
Why it matters
This development is significant for AI developers and cybersecurity professionals as it demonstrates how on-device AI can enhance user safety without compromising privacy. It also sets a precedent for integrating AI into messaging platforms while maintaining end-to-end encryption.
- On-device AI can enhance user safety without compromising privacy.
- Lightweight models like Gemma are essential for efficient threat detection on mobile devices.
- Strict governance and data privacy measures are critical for developing AI tools in sensitive domains.
- Collaboration between governments and AI providers can lead to innovative solutions for public safety.
The bigger picture
The model enables real-time detection of scam attempts, helping users avoid financial and credential exposure. It also provides a framework for integrating AI into messaging platforms while maintaining user privacy and security.
The collaboration between Google DeepMind and the Singapore Government highlights the potential of on-device AI to combat scam-related crimes while preserving user privacy. This project sets a new standard for integrating AI into messaging platforms with a focus on security and data protection.
Technical details
- The Gemma model is lightweight and optimized for mobile devices.
- A two-stage pipeline is used: a lightweight classifier for initial threat detection and a contextual evaluation with the fine-tuned model for confirmation.
- The model runs on-device, avoiding cloud-based data egress and preserving end-to-end encryption.
- Strict governance measures, such as the Sovereign Sandbox and anonymized datasets, were used to ensure data privacy.
FAQ
What is the main purpose of the new AI tool?
The tool is designed to detect scam conversations in private chats and protect citizens from financial and credential exposure.
How does the model ensure user privacy?
The model processes conversation data directly on the user's device, avoiding cloud-based data egress and preserving end-to-end encryption.
What is the two-stage pipeline used for?
The two-stage pipeline first uses a lightweight classifier to detect potential threats and then uses the fine-tuned Gemma model to confirm or discard threats.
What data was used to train the model?
The model was fine-tuned on anonymized datasets derived from scam reports and intelligence feeds, excluding private or personally identifiable information.
What is the significance of this collaboration?
This collaboration demonstrates how AI can be used to enhance user safety without compromising privacy, setting a precedent for integrating AI into messaging platforms.
This summary and analysis were generated with AI from the original article and may contain errors. Read the original for the full story.
Read the full story
Originally published by Google DeepMind on Oct 14, 2026.
Comments
Sign in to join the discussion.