The technology, announced on Wednesday at Singapore International Cyber Week, uses a customised version of Google’s Gemma model to recognise manipulation in conversations and warn users before they transfer money or disclose sensitive information.
The partners have not announced a public release date, named supported messaging applications or published independent accuracy figures. Their announcement describes a model that messaging platforms and operating system developers could integrate, rather than a scam detection application already available for general download.
The initiative forms part of the National AI Partnership between Google DeepMind and Singapore’s government. It addresses fraud schemes that unfold through extended conversations, rather than relying exclusively on suspicious links, keywords or individual messages.
Google DeepMind said the model had been adapted using examples of scams involving impersonation of banks, public institutions and prominent local figures. Its training also covered regional fraud patterns intended to make fraudulent exchanges appear credible to Singaporean users.
The system operates through two stages to limit the processing demands placed on smartphones. A lightweight classifier initially examines incoming messages for possible threats, consuming relatively little computing power. Messages considered suspicious then undergo a more detailed assessment using the fine tuned Gemma model.
During the second stage, the AI examines a broader section of the conversation to assess whether the exchange displays patterns of manipulation. This contextual approach is intended to distinguish ordinary conversations from attempts to build trust before requesting payments, passwords or other confidential details.
Where the model identifies potentially fraudulent behaviour, it generates a warning on the handset, encouraging the recipient to pause and exercise caution. The alert is designed to intervene before a person acts on a scammer’s instructions, although the organisations have not disclosed measured reductions in fraud losses from using the system.
A central feature is that the analysis takes place locally. Conversations do not need to be uploaded for cloud based examination, an important distinction for messaging services that use end to end encryption to prevent intermediaries from reading message contents.
Mark Chen, director of trust and safety at GovTech Singapore, said the collaboration sought to identify targeted local scams in real time while keeping personal communications private. The partners present local processing as a way of reconciling fraud prevention with confidentiality.
The privacy safeguards also extend to development. Google DeepMind said the model was fine tuned inside a government managed Google Cloud environment described as a sovereign sandbox, where training material remained isolated.
According to the company, private citizen conversations and personally identifiable information were excluded from training. Instead, developers used anonymised datasets derived from scam reports and intelligence feeds. Testing included synthetic conversations and local material covering different scam categories and changing impersonation methods.
Those measures describe the development process and proposed operation, rather than an independent assessment of the finished product. The partners have not released detailed public benchmarks showing false positive rates, missed scams, battery consumption or performance across different smartphone models.
Singapore has strong reasons to pursue additional protections. Google DeepMind said scams account for more than half of crimes reported in the country, while cumulative losses since 2019 exceed S$4 billion. It estimated that social engineering schemes account for as much as 80 per cent of scam losses.
Such schemes often depend on prolonged interaction, with perpetrators establishing apparent relationships or authority before introducing requests for money. Detecting suspicious intent across multiple messages therefore presents a different technical challenge from blocking known fraudulent websites or telephone numbers.
The announcement comes alongside separate Android security enhancements unveiled by Google, including optional cloud assisted analysis in Google Messages. That service uses a protected computing environment, whereas the Singapore collaboration specifically centres on keeping conversation analysis on the handset.
The distinction is significant because the two approaches involve different arrangements for processing private communications. The Singapore model is designed to assess message content without transmitting conversations externally, although its practical privacy protections will also depend on how individual applications implement the technology.
Google DeepMind said its model was optimised for smartphones commonly used in Singapore. It has not specified the minimum hardware requirements, participating mobile manufacturers or whether the technology will eventually support additional languages and markets.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
