Google DeepMind and the Singapore Government have unveiled a new AI tool designed to protect citizens from conversational scams without compromising privacy. The solution uses a lightweight, on-device Gemma model to detect manipulative conversation patterns in private chats, with all processing happening locally on the user's smartphone. Scams now account for over half of Singapore’s reported crimes and more than S$4 billion in losses since 2019, with up to 80% stemming from sophisticated social engineering tactics.
The model was fine-tuned on local scam conversations, including impersonations of banks, public institutions, and regional figures, to distinguish benign chatter from genuine manipulation. By operating entirely on-device, the tool avoids cloud data egress, making it compatible with end-to-end encrypted messaging platforms. A two-stage pipeline ensures efficiency: a lightweight classifier screens messages for potential threats, followed by contextual evaluation using the fine-tuned Gemma model to confirm or dismiss risks.
Real-time warnings appear directly on the device when manipulative intent is detected, prompting users to pause before sharing sensitive information. Developed under strict governance within GovTech’s Sovereign Sandbox on GCP, the model was trained exclusively on anonymised, non-personal data from scam reports and intelligence feeds, and stress-tested against synthetic dialogues and local corpora. Presented at Singapore International Cyber Week 2026, the initiative demonstrates how public-private collaboration can advance AI for public good, proving that safety and privacy can coexist through edge-based protection.
Source: Google DeepMind Blog · Summarized by HeadlinesBriefing