Lightweight Gemma models intercept scams locally, keeping chat data private on devices

Google DeepMind and GovTech Singapore announced a new artificial intelligence tool designed to protect citizens from scams. The solution relies on lightweight Gemma models that run entirely on the user's device. This approach allows the system to intercept conversational scams within private chats while ensuring zero data egress to the cloud.

How does the technology function?

The system utilizes Gemma models that are optimized for on-device execution. According to Google DeepMind, these models analyze conversations locally to identify scam patterns. Because the processing happens on the device, no chat data is uploaded to external servers. This architecture maintains user privacy while providing real-time detection of malicious interactions.

What is the operational scope?

The announcement highlights a partnership between Google DeepMind and the Singapore government agency GovTech. The focus is on protecting citizens against conversational scams. The tool operates within private messaging environments, targeting the specific threat of social engineering attacks delivered through chat. This marks a shift toward privacy-preserving security measures in public sector technology deployments.

Why is privacy preservation significant?

Traditional scam detection often requires sending data to the cloud for analysis. This new method eliminates that requirement by using on-device processing. Users retain control over their data, as chats remain private. The absence of cloud data egress reduces the attack surface for potential data breaches and aligns with stricter data protection standards.

What are the implications for security teams?

Security operations teams should note the trend toward edge-based AI security. This deployment demonstrates that effective threat detection can occur without compromising data privacy. Organizations may need to evaluate their own reliance on cloud-based analysis for threat intelligence. The success of this model could influence future procurement decisions for communication security tools.

Questions to ask your suppliers

Teams should verify whether vendors process sensitive data locally or send it to external servers for analysis. Check if the supplier provides clear documentation on data residency and egress policies. Confirm that the vendor’s security controls align with your organization’s privacy requirements. Discuss how updates to detection models are deployed without compromising user data. Ensure that the supplier has a defined incident response plan for potential data leaks. Inform legal and compliance teams of any changes in data handling procedures.

Background: Data residency

Data residency is the requirement that digital information remains stored within specific geographic boundaries. These rules exist to satisfy local laws or organizational policies. You must configure cloud storage to prevent data from moving across borders. This control limits where your backups and active records physically sit.

Read the full guide: What Is Data Residency: How Location Rules Protect Your Cloud

What to do and how to stay safe: Google DeepMind

• Review current communication security tools to assess where data processing occurs.
• Verify that vendors provide transparency regarding data egress and privacy controls.
• Monitor for new on-device security solutions that reduce cloud dependency.
• Update internal policies to reflect the use of privacy-preserving AI tools.

Step-by-step guide: Browser Security Settings for Small Business Teams

General security guidance from the TechSec Feed newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Does the AI tool send chat data to the cloud?

No, the Gemma models run on-device with zero cloud data egress, keeping chats private.

Who developed this scam detection tool?

Google DeepMind and GovTech Singapore collaborated to create the on-device AI tool.

What type of scams does it detect?

The tool is designed to intercept conversational scams within private chats.