New AI Tool to Detect Scams on Smartphones Developed by Google and Singapore's GovTech
A collaboration between Google and Singapore's GovTech has led to an innovative AI model that can identify scams in real-time on smartphones.
In a groundbreaking initiative, Google DeepMind has partnered with Singapore's Government Technology Agency, commonly known as GovTech, to create an on-device artificial intelligence model aimed at detecting scams within smartphone conversations. This novel approach focuses on privacy by ensuring that conversations are not sent to the cloud for analysis.
Addressing the Rising Threat of Scams
Presented during the Singapore International Cyber Week 2026, the new AI model is part of Google's broader National AI Partnership initiative with the Singaporean government. With scams accounting for over half of reported crimes in Singapore, the urgency for such technology is clear, especially given that losses related to scams have surpassed S$4 billion since 2019.
Recent trends indicate that many scammers have adopted social engineering tactics, which manipulate victims through convincing dialogues, leading to approximately 80% of all scams. Google DeepMind has identified this alarming shift, highlighting the need for enhanced detection methods.
How the AI Model Works
The AI system is based on Google’s lightweight Gemma model specifically tailored for this purpose. It has undergone rigorous fine-tuning to better understand local conversations, particularly those involving impersonations of banks, government institutions, and well-known local personalities.
To optimize battery life on smartphones, a two-stage detection process is employed. Initially, a lightweight classifier scans incoming messages for potential threats. In the event of a flagged message, the Gemma model performs a deeper analysis of a wider portion of the conversation to ascertain whether it is a legitimate threat or a false positive.
If the model detects patterns indicative of manipulation, it promptly issues a real-time alert to the user, advising them to be cautious before sharing any sensitive information or sending money.
Privacy at the Forefront
A notable advantage of this AI model is its operational framework, which processes all conversations directly on the user’s device. This means that the analysis happens entirely offline, maintaining the privacy of personal communications—a crucial element, especially in today’s digital landscape.
Mark Chen, the Director of Trust and Safety at GovTech, emphasized the importance of this project by stating, “Fighting scams shouldn’t mean sacrificing personal privacy.” This partnership aims to combat localized scams while ensuring that individual user privacy is uncompromised.
Next Steps for Deployment
Google DeepMind and GovTech have plans to roll out this innovative model gradually, integrating it into related applications developed in conjunction with the Singapore government. The implementation will only occur with users' consent, adhering strictly to the terms of the applications involved.
Both organizations have committed to ensuring that neither Google nor the Singapore government will have access to the conversations processed by the model. During its development, the teams worked within a controlled Google Cloud environment managed by GovTech. This isolation kept training data secure, utilizing anonymized datasets from various scam reports while deliberately excluding any personally identifiable information.
Additionally, extensive testing was conducted using both synthetic conversations and local datasets that reflected various scam incidents and the evolving tactics used by impersonators.
As of now, no official timeline for deployment or details about participating applications have been shared. However, Google DeepMind has indicated that they are investigating the feasibility of implementing similar technologies in other countries, suggesting that this innovative approach could expand its reach well beyond Singapore.
