Google DeepMind and GovTech Singapore Build On-Device Scam AI Using Gemma
Google DeepMind and Singapore's GovTech have launched an on-device scam detection model using a fine-tuned Gemma model to protect users while preserving message privacy.
Google DeepMind and the Government Technology Agency of Singapore (GovTech) have unveiled an on-device scam detection system designed to identify fraudulent communications directly on smartphones. Announced at Singapore International Cyber Week (SICW 2026), the project was developed under the National AI Partnership with the Singapore Government. By running model inference locally on user hardware, the system addresses a surge in conversational fraud without compromising user privacy or bypassing end-to-end encryption.
In Singapore, scams account for more than half of all reported crimes and have caused over S$4 billion in financial losses since 2019. Modern fraudulent operations have largely moved away from crude phishing schemes toward complex social engineering tactics, where attackers build rapport across extended exchanges before requesting funds or credentials. These manipulative conversations are now responsible for up to 80% of all scam-related losses. Because standard messaging services increasingly protect communications with end-to-end encryption, central servers cannot read message text to filter threats, creating an urgent need for secure, client-side moderation.
Two-stage edge detection architecture
To balance protective scanning with edge power constraints, Google DeepMind contributors JJ Geewax and Miguel de Andrés detailed a two-stage evaluation pipeline optimized for mobile processors:
- Stage 1 (Lightweight Screening): A low-power classifier monitors incoming text to catch early indicators of manipulative patterns while minimizing battery drain.
- Stage 2 (Contextual Verification): When the initial classifier flags an anomaly, a fine-tuned Gemma model evaluates a broader context window of the exchange to verify genuine manipulation or discard false positives.
If the system confirms fraudulent intent, it issues an immediate on-device advisory warning the user to exercise caution before transferring funds or sharing credentials. Running this analysis directly on the smartphone ensures that message content never leaves the device for threat inspection.
Localized fine-tuning and strict data governance
Standard language models frequently fail to recognize localized deception techniques, such as regional colloquialisms or impersonations of local public agencies, banks, and community figures. To overcome this limitation, the team fine-tuned Gemma on regional scam typologies, validating its performance against synthetic conversations and local text corpora.
Development occurred under strict governance protocols to maintain data confidentiality. All model fine-tuning took place within GovTech's managed Google Cloud Platform (GCP) sovereign sandbox. The training process explicitly excluded private citizen information or personally identifiable data, relying entirely on anonymized scam reports and threat intelligence feeds.
What it means for developers
This deployment provides a clear structural blueprint for developers building security tools, chat clients, and privacy-sensitive mobile applications. As end-to-end encryption becomes standard across modern software, cloud-based inspection is increasingly impractical for user moderation. Instead, software engineers can adopt client-side multi-stage architectures, using lightweight filters as gatekeepers for small language models (SLMs) running locally on mobile hardware.
For engineering teams experimenting with localized threat classification, conversational safety guardrails, or hybrid edge-cloud pipelines, testing model capabilities across multiple architectures is an essential step. Developers can try top AI models cheaply through one API at https://apixoai.online to benchmark latency, prompt performance, and classification accuracy across different frontier providers.
The DeepMind and GovTech implementation demonstrates that defensive AI does not need to compromise privacy to remain effective. By coupling small, fine-tuned open models like Gemma with power-efficient edge pipelines, developers and operating system maintainers can protect end users against social engineering in real time.
Source: Google DeepMind and Singapore Government announce new AI tool to protect citizens against scams — Google DeepMind. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
