Google DeepMind unveils on-device scam tool in Singapore
Google DeepMind and Singapore’s Government Technology Agency have developed an on-device scam detection model designed to identify manipulative conversations without sending chat data to the cloud. The lightweight system was presented at Singapore International Cyber Week 2026.
Radwan AbdulhadiEditor and technology writer
AI model targets conversational scams
Google DeepMind and the Government Technology Agency of Singapore, known as GovTech, have developed an on-device AI model intended to warn people about scams conducted through private conversations. The collaboration addresses fraud that relies on social engineering, in which attackers build convincing exchanges before asking victims to transfer money or reveal credentials. Google DeepMind described the work in its announcement on an on-device scam detection model.
According to Google DeepMind, scams account for more than half of Singapore’s reported crimes and have caused over S$4 billion in losses since 2019. The company said scams involving complex conversational manipulation now account for as much as 80% of all losses, as criminal tactics have moved beyond simpler phishing attempts.
Designed for local scam patterns
The system uses a lightweight Gemma model that was fine-tuned on local scam conversations. The training material covered impersonations of public institutions, banks and local figures, as well as regional schemes that were trending during development. That local focus is intended to help the model distinguish ordinary social exchanges from conversations designed to manipulate a recipient.
The project was carried out as part of the National AI Partnership between Google DeepMind and the Singapore Government. The partnership enabled GovTech and Google DeepMind to work on a model optimised for typical smartphones in Singapore, while addressing the need to protect people from fraud without exposing the contents of private communications.
Analysis stays on the device
Rather than sending conversation data to a remote service for inspection, the model processes messages directly on the user’s device. Google DeepMind said this approach is particularly relevant because many messaging platforms use end-to-end encryption, which prevents conventional server-side systems from examining the content of private chats.
When the model identifies manipulative conversational patterns, it can trigger a real-time warning on the smartphone. The warning advises the user to pause and exercise caution before sharing sensitive information or sending funds. Local analysis is also intended to allow people to respond before their financial information or account credentials are exposed.
The approach is presented as a way to combine fraud protection with private communications. Google DeepMind said bringing threat analysis to the device means personal conversations can remain on the smartphone while the system evaluates signals associated with a scam.
Two-stage pipeline limits battery use
Running a complex classifier on a smartphone creates practical challenges, particularly around battery consumption. To address that, the system uses a two-stage pipeline. The first stage is a lightweight classifier that screens incoming messages using minimal power and identifies conversations that may require closer examination.
The second stage uses the fine-tuned Gemma model to evaluate a larger snapshot of the conversation. That contextual analysis is intended to confirm a potential threat or discard a false alarm, rather than applying the more computationally demanding assessment to every message.
Governance and testing measures
Google DeepMind said the fine-tuning process took place inside a managed Google Cloud environment maintained by GovTech, which it called a sovereign sandbox. The arrangement kept the training data isolated during development.
The company said private or personally identifiable citizen data was excluded from training. Instead, the model used anonymised datasets derived from scam reports and intelligence feeds. It was also stress-tested with synthetic dialogues and local corpora covering different scam categories and changing impersonation tactics.
The collaboration was presented at Singapore International Cyber Week 2026 as an example of how messaging platforms and operating-system vendors could integrate lightweight models locally. Google DeepMind said such integrations could help services counter criminal activity while preserving end-to-end encryption and avoiding exposure of private chat data.
Sources
- Google DeepMindGoogle DeepMind and Singapore Government announce new AI tool to protect citizens against scams
Continue chronologically
Related entity coverage
Related stories
Anthropic Warns Self-Replicating AI Could Target IoT Devices
Anthropic warns that small AI models could soon self-replicate across consumer IoT devices, forming distributed botnets without data centers.
Google DeepMind Details Gemini 4 Argon Safety and Security
Google DeepMind outlines safety, security, and mitigation measures for Gemini 4 Argon, addressing misuse, alignment, and real-time safeguards.
Google DeepMind Unveils Antigravity Cheminformatics
Google DeepMind introduces Google Antigravity with new Cheminformatics Skills to accelerate drug discovery workflows and design novel macrocycles.
Google DeepMind Unveils Genie 4 World Simulation Model
Google DeepMind introduces Genie 4, a new real-time interactive world simulation model featuring native audio, advanced physics, and Gemini integration.
