Google DeepMind has partnered with Singapore’s Government Technology Agency (GovTech Singapore) to develop a privacy-preserving, on-device AI model to counter localized conversational scams. Built by fine-tuning Google’s lightweight Gemma model on local fraud data, the system detects complex social engineering techniques such as impersonations of public institutions and local financial scams.

Unlike traditional cloud-based threat detection systems, the fine-tuned model processes conversations locally on smartphones with user consent. Operating entirely on-device ensures end-to-end encryption and user privacy, as neither Google nor the Singaporean government can access message content.

Unveiled at Singapore International Cyber Week 2026, the two-stage pipeline triggers real-time alerts when manipulative patterns appear. The initiative serves as a technical framework for governments and messaging platforms seeking to deploy proactive anti-fraud measures without compromising data sovereignty.

Why it matters

  • Demonstrates commercial viability of running lightweight open models (Gemma) entirely on-device for latency-sensitive applications.

  • Establishes a blueprint for privacy-preserving AI models that balance localized threat detection with end-to-end encryption.

  • Highlights growing public-private partnerships between major model providers and sovereign governments.

Source: deepmind.google