Google DeepMind and Singapore GovTech Develop On-Device AI for Scam Detection
Fine-tuned Gemma models deliver real-time local fraud alerts while maintaining complete message privacy.
1Singapore Confronts Escalating Fraud
Scams represent over half of Singapore's reported crimes, causing upwards of S$4 billion in cumulative losses since 2019. Threat tactics have evolved beyond rudimentary phishing toward sophisticated social engineering schemes. These conversational manipulation attacks now account for approximately 80% of all recorded monetary fraud across the island nation.
2Strategic National AI Partnership
To counter escalating threat vectors, Google DeepMind partnered directly with the Government Technology Agency of Singapore. Operating under a broader National AI Partnership, the collaborative effort engineered a dedicated, privacy-focused fraud detection system. The initiative aims to deploy proactive defensive intelligence into applications widely utilized by everyday citizens.
3Localized Gemma Model Fine-Tuning
Generic AI classifiers frequently fail against regional linguistic nuances and hyper-local pretexting. Engineers addressed this by fine-tuning Google's lightweight open Gemma model using genuine local scam transcripts. Training data spanned fraudulent impersonations of Singaporean public agencies, local commercial banks, public figures, and trending regional financial scams.
4Zero-Cloud Edge Architecture
Traditional anti-fraud systems rely on server-side telemetries and cloud text scanning, raising severe data exposure risks. Conversely, DeepMind's fine-tuned model executes analysis entirely on-device. This localized inference mechanism rapidly detects manipulative intent across conversation flows before users disclose critical financial details, personal credentials, or authorize unauthorized bank transfers.
5Guaranteed User Data Isolation
The system mandates strict user consent before inspecting message threads within host client applications. Because text evaluation takes place strictly inside edge silicon, neither Google nor Singaporean authorities obtain access to private conversations. This isolated processing framework preserves end-to-end communication secrecy without sacrificing real-time automated fraud threat mitigation.
6Real-Time Manipulative Intent Alerts
Whenever the edge-based Gemma model identifies conversational manipulation markers, it immediately issues a native on-device warning. This intervention alerts the user to pause, evaluate the context, and avoid transferring sensitive assets. Immediate contextual feedback breaks the emotional coercion cycle commonly leveraged by sophisticated professional fraud operators.
7Two-Stage Battery Efficient Pipeline
Continuous background classification on modern mobile hardware presents serious energy and thermal overhead challenges. The engineering team deployed a specialized two-stage execution pipeline to minimize hardware drain while preserving responsiveness. This optimization ensures continuous protection during messaging sessions without noticeably degrading mobile battery life or overall device performance.
8Global Blueprint For Encryption Safety
Unveiled at Singapore International Cyber Week 2026, the project demonstrates scalable defense alongside end-to-end encryption. The deployment establishes an actionable blueprint for sovereign governments and frontier AI labs. DeepMind intends to explore similar defensive edge deployments across foreign jurisdictions confronting comparable conversational social engineering threats.