Google, GovTech develop on-device AI model to detect scams in Singapore

Google DeepMind and GovTech Singapore are developing an AI model designed to detect scam conversations directly on Android devices in real time, as part of a wider collaboration between Google and the Singapore government on cybersecurity and anti-scam protection.

Unveiled at Singapore International Cyber Week (SICW) 2026, the privacy-preserving model uses Google's lightweight Gemma models, fine-tuned on local scam narratives including impersonations of public institutions, banks and local figures, as well as emerging regional schemes.

The model is designed to distinguish everyday conversations from the manipulation tactics commonly used in social engineering scams, which have increasingly moved beyond straightforward phishing attempts into prolonged and convincing conversations.

Don't miss: Google names new MD for SEA and South Asia frontier markets

According to Google, scams including investment fraud, fake job offers and impersonation of public officials make up more than half of reported crimes in Singapore and have accounted for over SG$4 billion in losses since 2019. Social engineering scams account for up to 80% of scam losses.

Rather than sending conversations to external cloud servers for analysis, the model processes them locally on a user's device. When integrated into applications, it will analyse conversational patterns and flag potentially manipulative behaviour, triggering a warning that prompts users to exercise caution before sharing sensitive information or transferring funds.

The model uses a two-stage system aimed at balancing scam detection with the computing and battery constraints of smartphones. A lightweight classifier first screens incoming messages for potential threats using minimal power, before the fine-tuned Gemma model analyses a larger portion of the conversation to determine whether a threat is present or dismiss a false alarm.

Google said the rollout will be phased and developed with the Singapore government. Once deployed, the model will only be able to access on-device conversations with users' consent and in accordance with the terms of the application in which it is integrated.

In addition, because the analysis takes place entirely on-device, Google and the Singapore government will not be able to access users' conversations through the model's deployment, according to the company.

The model was fine-tuned within a GovTech-managed Google Cloud environment, with training data kept isolated. Google said private and personally identifiable information from citizens was excluded from training, with anonymised datasets derived from scam reports and intelligence feeds used instead.

It was also tested against synthetic conversations and local datasets spanning different scam categories and evolving impersonation tactics.

Beyond scam detection

Alongside the scam detection project, Singapore is joining Google's Fairwind Programme, which gives cyber defenders access to cybersecurity tools including its Gemini 4 Argon model and CodeMender harness.

The programme is designed to help organisations identify, verify and patch vulnerabilities using AI, with Google saying the tools can generate verified, deployment-ready patches within an organisation's secure cloud environment.

The two initiatives build on Google's existing cybersecurity work with Singapore, including a refreshed 2025 memorandum of understanding with the Cyber Security Agency of Singapore (CSA) and an AI Agents Sandbox aimed at studying how computer-use agents operate in real-world environments.

On the consumer front, Google said enhanced fraud protection under Google Play Protect has blocked more than 3.3 million attempts to install potentially malicious applications across over 770,000 devices, while preventing 68,000 unique apps from potentially being used for financial fraud and scams.

Google has also been working on scam literacy through "Be scam ready", an interactive game designed to teach users how to recognise manipulation tactics. According to Google, studies found that 10 minutes of gameplay was associated with an estimated 20% lower overall risk of falling for a scam. The company is now rolling out "Be scam ready 2.0", which halves the playtime while aiming to maintain the same level of protection.

“Fighting scams shouldn't mean sacrificing personal privacy. Working with Google DeepMind, we're bringing advanced AI protection directly onto citizens' devices. This allows us to catch highly targeted, local scam attempts in real time while keeping personal communications private,” said Mark Chen, director, trust and safety, GovTech Singapore.

In tandem, JJ Geewax, director, applied AI, Google DeepMind said, “As scam tactics evolve, protecting people requires new, more advanced, more intelligent approaches. Pairing Google DeepMind’s research with GovTech Singapore’s local domain and data expertise means we can confront this urgent public safety challenge head-on."

“This initiative proves that digital safety and personal privacy can coexist on-device, offering a blueprint for how governments and model providers can collaborate to build technology to keep people safe,” he added.

The initiatives follow Google’s expansion of its AI footprint in Singapore earlier this year, spanning healthcare, enterprise innovation, talent development and online safety. Singapore has served as Google’s Asia Pacific headquarters since 2007, with the company investing US$5 billion in technical infrastructure and establishing a Google DeepMind research lab in the country.

Related articles: 
OpenAI, Anthropic and Google reportedly explore joint AI standards body   
Google Maps gets chatty with Gemini-powered Ask Maps in Singapore 
Google’s Sapna Chadha takes on expanded APAC ads remit