Google DeepMind and Singapore’s Government Technology Agency have developed an artificial intelligence model that detects manipulative scam conversations directly on smartphones. Announced on 14 October during Singapore International Cyber Week, the system is designed to issue real-time warnings before users transfer money, disclose credentials or share other sensitive information. Conversation data remains on the device rather than being transmitted to an external cloud service.
The project uses a lightweight version of Google’s Gemma model fine-tuned on Singapore-specific scam patterns. Its training material covers impersonation of government institutions, banks and public figures, as well as other social-engineering techniques seen in regional fraud campaigns. Private or personally identifiable citizen data was excluded from training; the partners used anonymised material derived from scam reports and intelligence feeds.
A two-stage architecture is intended to balance detection capability with the processing and battery constraints of consumer smartphones. The first stage uses a low-power classifier to screen incoming messages. Conversations showing potentially suspicious characteristics move to a second stage, where the fine-tuned Gemma model analyses a broader sequence of messages for signs of manipulation. If the system identifies a threatening pattern, it generates an on-device warning advising the user to pause before sharing information or funds.
The approach is designed for messaging environments protected by end-to-end encryption, where a platform operator cannot inspect private conversations on a central server. By moving the detection process onto the handset, the partners aim to preserve that encryption boundary while identifying behavioural patterns that may not be apparent from an individual message or link.
Development and fine-tuning took place inside a GovTech-managed Google Cloud environment. The model was stress-tested against synthetic conversations and local language patterns covering multiple scam categories. It is optimised for smartphones commonly used in Singapore, although a public release date and details of any integration into Android or specific messaging platforms have not been disclosed.
Scams account for more than half of reported crimes in Singapore and have caused over S$4 billion in losses since 2019. Social-engineering schemes, in which attackers build a sustained and apparently credible conversation before requesting money or credentials, now account for as much as 80% of scam losses. The model extends Singapore’s existing use of on-device screening, ScamShield intelligence and automated website analysis into multi-turn conversational detection.
