GovTech Singapore and Google DeepMind have developed an on-device scam detection system based on a fine-tuned Gemma AI model, presenting the architecture at SICW 2026.

The project forms part of the National AI Partnership with the Singapore Government. Fraud schemes have caused financial losses exceeding S$4 billion in Singapore since 2019, representing more than half of all reported crimes. Complex social engineering conversations, rather than automated mass-phishing links, now generate up to 80 percent of those recorded losses.

Mark Chen, Director of Trust and Safety at GovTech Singapore, comments: “Fighting scams shouldn’t mean sacrificing personal privacy.

“Working with Google DeepMind, we’re bringing advanced AI protection directly onto citizens’ devices. This allows us to catch highly targeted, local scam attempts in real time while keeping personal communications private.”

Two-stage pipeline runs on mobile hardware

Battery drain presents a persistent technical obstacle when running inference tasks on smartphones. To manage resource limits, the engineering teams established a tiered screening architecture.

Stage 1 runs a compact, low-power binary classifier directly on inbound messages. This initial filter monitors incoming text continuously, screening for early markers of manipulative behaviour while consuming minimal electrical power. When this stage flags potential deception, the system escalates the interaction to the secondary layer.

Stage 2 initiates an in-context evaluation using the fine-tuned Gemma model. The system analyses an expanded snapshot of the surrounding conversation to verify malicious intent or dismiss false alerts. If the model confirms manipulative dialogue, it issues an immediate warning on the handset, prompting the user to pause before providing credentials or transferring capital.

Sovereign cloud training isolates local data

Model customisation took place within GovTech Singapore’s managed Google Cloud Platform environment, configured as an isolated Sovereign Sandbox. This deployment kept developmental weights separate from outside cloud infrastructure.

The engineering teams deliberately 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. These datasets exposed the model to specific regional tactics, such as the impersonation of local public officials, domestic banks, and regional investment opportunities.

Consent controls govern edge deployment

The software operates locally rather than routing user conversations to centralised cloud servers for inspection. This architectural choice aligns with end-to-end encryption practices on messaging platforms.

Under the planned rollout framework, the detection pipeline accesses message streams only after receiving explicit user consent in accordance with application terms. Neither Google nor the Singapore Government receives access to citizen conversation data during model execution.

GovTech Singapore plans a phased rollout across public applications, with the technical framework serving as a reference implementation for other jurisdictions exploring edge-based fraud prevention.

See also: AISI outlines priorities to tackle frontier AI evaluation gaps

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including IoT Tech Expo and the Cyber Security & Cloud Expo. Click here for more information.

AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.