GovTech Singapore and Google DeepMind have developed an on-device AI scam detection system powered by a fine-tuned Gemma model to identify suspicious conversations while keeping personal messages on users’ devices. The technical architecture was presented at Singapore International Cyber Week (SICW) 2026 as part of the National AI Partnership with the Singapore Government.
The initiative addresses Singapore’s growing fraud problem. According to the project information, scams have caused more than S$4 billion in financial losses since 2019 and account for over half of reported crimes. Complex social engineering attacks, which manipulate victims through seemingly legitimate conversations, now account for up to 80% of recorded scam losses.
Unlike conventional fraud prevention systems that primarily focus on malicious links or suspicious websites, the new AI-powered approach is designed to identify deceptive language and manipulative behaviour within conversations.
Mark Chen, Director of Trust and Safety at GovTech Singapore, emphasised the importance of balancing effective fraud prevention with data protection.
“Fighting scams shouldn’t mean sacrificing personal privacy,” Chen said, explaining that on-device AI can help identify targeted scams in real time without exposing private communications to centralised inspection.
How the Two-Stage Gemma AI Scam Detection System Works
Running artificial intelligence continuously on smartphones presents several challenges, particularly battery consumption, processing capacity and response time. To address these limitations, GovTech Singapore and Google DeepMind designed a two-stage detection pipeline that combines lightweight screening with more advanced AI analysis.
Stage 1: Lightweight AI Filters Suspicious Messages
The first stage uses a compact, low-power binary classifier that runs directly on the mobile device. It continuously screens incoming text for early indicators of potentially manipulative behaviour while minimising battery usage.
When the classifier identifies a potentially suspicious interaction, it forwards the relevant content to the second stage for further evaluation.
This initial screening process helps reduce the computational demands of running a more sophisticated language model across every incoming message.
Stage 2: Fine-Tuned Gemma Model Analyses Conversations
The second stage activates a fine-tuned Gemma AI model to examine a broader snapshot of the conversation. By evaluating additional context, the system aims to distinguish genuine scam attempts from legitimate messages that might otherwise trigger false alarms.
If the model identifies persuasive or deceptive dialogue consistent with a scam, it generates a warning directly on the smartphone. The alert encourages users to reconsider potentially risky actions, such as sharing account credentials or transferring money.
Combining lightweight classification with contextual analysis enables the system to balance detection accuracy, device performance and energy efficiency.
Secure Cloud Training Supports Regional Scam Detection
Although scam detection takes place on the device, model customisation requires a controlled training environment. The engineering teams used a managed Google Cloud Platform environment operated by GovTech Singapore and configured as an isolated Sovereign Sandbox.
This setup was designed to separate development activities and model weights from external cloud infrastructure.
The project also excluded private citizen communications and personally identifiable information from its training datasets. Instead, the model was trained using anonymised fraud reports, official intelligence records, regional dialect data and synthetically generated chat conversations.
These datasets help the AI recognise scam tactics relevant to Singapore and the wider region, including impersonation of government officials, domestic banks and fraudulent investment opportunities.
Incorporating regional language patterns and locally relevant fraud scenarios is important because scammers often tailor their messages to the people and communities they target.
On-Device AI Prioritises Privacy and User Consent
A central feature of the system is its ability to analyse suspicious conversations locally rather than routinely transmitting message content to centralised cloud servers for inspection.
This approach supports privacy-conscious fraud prevention and can complement messaging services that use end-to-end encryption. However, the actual compatibility and implementation depend on the permissions and technical design of each application.
Under the planned deployment framework, the detection system will access message streams only after users provide explicit consent in accordance with the relevant application terms.
GovTech Singapore and Google DeepMind have stated that neither organisation receives access to citizens’ conversation data during on-device model execution.
Keeping inference on the handset reduces the need to transmit sensitive conversational content for analysis. It also gives users a way to receive scam warnings without requiring every message to be examined by a centralised service.
Singapore Plans a Phased Rollout of AI Scam Protection
GovTech Singapore plans to introduce the detection framework through a phased rollout across public applications. The technical architecture is also intended to serve as a reference implementation for other jurisdictions exploring privacy-preserving, edge-based fraud prevention.
The project highlights a broader shift in cybersecurity towards AI systems that can identify emerging threats directly on personal devices. Rather than relying exclusively on databases of known malicious links or centralised monitoring, contextual AI models can assess the language and intent behind suspicious interactions.
Nevertheless, on-device scam detection should complement, rather than replace, established security practices. Users should continue verifying unexpected requests for money, protecting authentication credentials and contacting financial institutions through official channels when suspicious activity occurs.
As social engineering scams become more sophisticated, combining locally executed AI, contextual analysis and explicit user consent could offer a practical direction for mobile fraud prevention.
The collaboration between GovTech Singapore and Google DeepMind demonstrates how the Gemma AI model can be adapted for on-device scam detection, combining lightweight message screening with deeper conversational analysis.
By keeping inference on smartphones and using controlled training processes, the initiative aims to improve the identification of targeted scams while reducing unnecessary exposure of personal communications.
The planned rollout will help determine how effectively this approach can support Singapore’s wider anti-scam efforts and provide a model for other countries seeking to strengthen digital security without compromising user privacy.