Google DeepMind said in a blog post on Oct. 14 that it has developed a scam detection AI model that runs on smartphones with Singapore's Government Technology Agency, GovTech Singapore.

The collaboration is part of a national AI partnership between Google DeepMind and the Singapore government.

Google DeepMind fine-tuned its lightweight Gemma model with local scam conversation data. It incorporated training on tactics ranging from impersonation of public agencies, banks and local celebrities to methods that are trending locally. That allows it to distinguish everyday conversations from actual manipulation attempts, the company said.

The model analyses conversations on the device rather than on cloud servers. When it detects manipulation patterns, it displays a real-time warning urging users to pause before sending money or sharing sensitive information. The model can access conversations only with user consent. Neither the Singapore government nor Google can view user conversations.

The model developed by the two sides uses a two-stage structure to improve battery efficiency. In the first stage, a lightweight classifier filters incoming messages. In the second, the fine-tuned Gemma analyses the conversation more broadly and makes a final determination on whether it is a threat or a false alarm.

The fine-tuning was carried out in a Google Cloud environment managed by GovTech Singapore. The training data excluded personally identifiable information and used only anonymised data extracted from scam reports and information feeds. The model will be rolled out into apps in phases after consultation with the Singapore government.

Mark Chen (마크 첸), director of Trust and Safety at GovTech Singapore, said, "Fighting scams should not mean sacrificing privacy," adding, "We can catch real-time scam attempts targeting locals while keeping private conversations confidential."