Google and Singapore Put Scam Detection on Phones
Google DeepMind and Singapore’s GovTech unveiled an on-device Gemma system that flags localized scam conversations without sending chats to the cloud.
A local defense against social engineering
Google DeepMind and Singapore’s Government Technology Agency have introduced an on-device AI system designed to detect scam conversations before users transfer money or reveal credentials. The collaboration targets a problem that has become central to Singapore’s crime statistics: scams account for more than half of reported crimes, while cumulative losses since 2019 have exceeded S$4 billion.
The system uses a lightweight screening classifier to examine incoming messages and escalate suspicious exchanges to a fine-tuned Gemma model. The second stage analyzes a broader slice of the conversation, looking for manipulation patterns such as impersonation of banks, public agencies or recognizable local figures. When the system finds a likely scam, it generates a warning directly on the phone.
Why the architecture matters
The model is designed to operate on typical smartphones rather than rely on continuous server-side inspection. That is important in encrypted messaging environments, where platforms may not be able to read conversations in the cloud. Processing locally can also reduce response time and limit the amount of sensitive material leaving the device.
Singapore says the model was trained in a GovTech-managed environment using anonymized scam reports and intelligence feeds, not private citizen conversations. The partners also stress-tested it with synthetic dialogues and local scam patterns. The system was presented during Singapore International Cyber Week and is framed as a model for cooperation between governments, platforms and model providers.
Why it matters
This is more than another fraud classifier: it tests whether small, specialized models can become a practical safety layer inside consumer devices. The unresolved questions are false positives, battery use, language coverage and who controls future updates. If those issues are solved, on-device models could let messaging services add protection without weakening end-to-end privacy.