Digital Guardian: GovTech Singapore and Google DeepMind Unveil On-Device AI Scam Detector

GovTech Singapore and Google DeepMind have unveiled a groundbreaking on-device scam detection system, leveraging a fine-tuned Gemma AI model. This innovative architecture was presented at SICW 2026, highlighting a crucial step in the National AI Partnership with the Singapore Government. The initiative addresses the escalating problem of financial losses due to fraud in Singapore, which has surpassed S$4 billion since 2019 and accounts for over half of all reported crimes. Notably, complex social engineering conversations are now responsible for up to 80 percent of these recorded losses, rather than traditional mass-phishing links.

Mark Chen, Director of Trust and Safety at GovTech Singapore, underscored the project's commitment to privacy, stating, “Fighting scams shouldn’t mean sacrificing personal privacy.” This ethos is central to the collaboration with Google DeepMind, aiming to deliver advanced AI protection directly onto citizens’ devices. This approach enables the system to detect highly targeted, localized scam attempts in real-time while rigorously preserving the privacy of personal communications.

A two-stage pipeline is engineered to run efficiently on mobile hardware, carefully managing the persistent technical obstacle of battery drain on smartphones during inference tasks. The first stage involves a compact, low-power binary classifier that operates directly on inbound messages. This initial filter continuously monitors incoming text for early indicators of manipulative behavior, consuming minimal electrical power to ensure continuous protection without significant battery impact.

Should the first stage identify potential deception, the system escalates the interaction to a secondary layer. Stage 2 then initiates an in-context evaluation using the sophisticated fine-tuned Gemma model. This stage analyzes an expanded snapshot of the surrounding conversation to thoroughly verify malicious intent or dismiss any false alerts. If the model confirms manipulative dialogue, it issues an immediate warning on the user's handset, prompting them to pause and reconsider before providing sensitive credentials or transferring capital.

Crucially, model customization took place within GovTech Singapore’s managed Google Cloud Platform environment, configured as an isolated Sovereign Sandbox. This setup ensured that developmental weights remained separate from outside cloud infrastructure, maintaining data sovereignty. The engineering teams meticulously excluded private citizen communications and personally identifiable records from the training data. Instead, the model learned from anonymised fraud reports, official intelligence records, regional dialect corpora, and synthetically constructed chat records. This specialized training exposed the model to specific regional tactics, such as the impersonation of local public officials, domestic banks, and unique regional investment opportunities, enhancing its effectiveness against local threats.

The software’s operational design emphasizes consent controls governing edge deployment. It functions locally on the device, meaning user conversations are not routed to centralized cloud servers for inspection. This architectural choice is in strict alignment with end-to-end encryption practices prevalent on modern messaging platforms. Under the planned rollout framework, the detection pipeline will only access message streams after receiving explicit user consent, fully adhering to application terms. Neither Google nor the Singapore Government will have access to citizen conversation data during the model’s execution, reinforcing user privacy. GovTech Singapore plans a phased rollout across public applications, with this technical framework serving as a reference implementation for other jurisdictions exploring similar edge-based fraud prevention solutions.