Google DeepMind, GovTech Unveil On-Device Scam Detection Tool for Singapore
Google DeepMind and Singapore's Government Technology Agency (GovTech) on October 14, 2026, announced an on-device scam detection tool built to catch conversational scams inside private messaging apps, issuing real-time warnings without sending user data to the cloud. Minister for Digital Development and Information Josephine Teo unveiled the strategic collaboration at Singapore International Cyber Week (SICW).
The system is designed to catch manipulation as a conversation unfolds, warning users before they transfer money, disclose credentials or share other sensitive information. Because the analysis runs entirely on the handset, the partners say it can work even inside end-to-end encrypted messaging environments where platform operators cannot inspect private chats on a central server.
How the On-Device Scam Detection Model Works
The tool uses a lightweight version of Google's Gemma model that was fine-tuned on Singapore-specific scam patterns, per a Google DeepMind blog post written by JJ Geewax and Miguel de Andrés. Training material covered impersonations of public institutions, banks and local figures, as well as trending regional fraud schemes, so the model can distinguish harmless social chatter from genuine manipulation.
To keep the classifier practical on consumer smartphones, the teams built a two-stage pipeline. A low-power classifier first screens inbound messages using minimal energy, and conversations that look suspicious are escalated to the fine-tuned Gemma model, which evaluates a larger snapshot of the exchange to confirm a threat or discard a false alarm. Battery efficiency was a central design concern, the announcement said.
If the system spots signs of manipulation in a chat, it triggers an on-screen alert telling the user to hold off before sending money or sensitive details. The goal is early intervention, ideally before funds move, and the resulting model is optimized to run on typical smartphones used in Singapore.
Privacy Design and Governance
The partners framed privacy as the project's defining constraint. Conventional threat detection scans messages on cloud servers, but many messaging platforms are end-to-end encrypted and cannot inspect private conversations, the announcement explained. Running the analysis locally keeps conversation data on the device, with no transmission to an external cloud service.
"Fighting scams shouldn't mean sacrificing personal privacy," said Mark Chen, Director of Trust and Safety at GovTech Singapore.
The announcement describes governance measures it labels Sovereign Sandbox, No Citizen Data and Stress-Testing. Fine-tuning took place inside a GovTech-managed and maintained Google Cloud Platform environment that kept training data isolated, private or personally identifiable information was strictly excluded in favor of anonymised datasets drawn from scam reports and intelligence feeds, and the model was validated using synthetic conversations and local language samples spanning multiple fraud types and shifting impersonation methods.
Why Singapore Is Targeting Conversational Scams
According to the announcement, more than half of all reported crimes in Singapore are scams, with losses exceeding S$4 billion since 2019. The partners say tactics have evolved beyond basic phishing into sophisticated social engineering, where attackers construct believable dialogues to manipulate victims before requesting payment, and such schemes now account for up to 80% of all losses.
Teo positioned the collaboration within a wider effort to combat AI-powered scams, observing that the challenge has intensified as fraudsters adopt increasingly advanced AI tools. Generic models often struggle with localised deception techniques, the announcement said, which is why the Gemma model was tuned on Singapore-specific material.
The tool adds to an existing stack of government defenses. SATIS, the Scam Analytics and Tactical Intervention System, operates for the Singapore Police Force and takes down roughly 20,000 malicious websites each month, while GovTech's rMSE system scans over 600,000 websites daily and ScamShield provides another layer of protection. The new model extends that apparatus into multi-turn conversational detection.
Rollout Questions and Platform Integration
Teo did not give a launch date for the Gemma-based tool and did not disclose performance metrics like detection or false-positive rates. The model is optimised for smartphones commonly used in Singapore, but the partners have not said whether it will be built into Android or particular messaging services, according to CXO Digitalpulse.
The announcement says messaging platforms and operating system vendors can directly integrate the lightweight models to safeguard users locally, and it presents the project as a template for balancing user protection with end-to-end encryption. A detection model only helps if it is built into the apps people actually use, which typically requires cooperation from messaging platforms and device makers.
The work was delivered under Google's National AI Partnership with the Singapore Government, which Google DeepMind announced on May 20, 2026, introducing new country programmes under its National Partnerships for AI initiative following the opening of a Singapore research lab the previous year. The collaboration was presented at Singapore International Cyber Week 2026, where the opening address was delivered on October 13 by Senior Minister and Minister for Home Affairs K Shanmugam, according to the address published by the Ministry of Home Affairs.
Accuracy and Adaptation Challenges
The project faces open questions beyond timing. A detector that raises false alarms too frequently risks being tuned out by users, while one that overlooks real threats provides little meaningful defense, so accuracy will determine whether warnings change user behavior.
Fraudsters also shift their methods rapidly, and Teo herself noted that AI is complicating efforts to stop them. A model trained on current scam scripts will require ongoing retraining to remain effective against new ones, a maintenance burden the partners have not yet detailed publicly.
Conclusion
Google DeepMind and GovTech have built a lightweight, Gemma-based model that flags manipulative scam conversations on the device itself, fine-tuned on Singapore-specific schemes and governed by rules that keep citizen data out of training. Announced at Singapore International Cyber Week on October 14, 2026, the on-device scam detection system is optimized for common local smartphones and designed for end-to-end encrypted messaging. The next concrete step is integration: the partners say messaging platforms and operating system vendors can embed the models directly, though no launch date or performance figures have been disclosed.
Frequently Asked Questions
What did Google DeepMind and GovTech Singapore announce on October 14, 2026?
They announced an on-device AI model, built on a lightweight version of Google's Gemma, that detects manipulative scam conversations in private messaging apps and warns users in real time. Minister Josephine Teo unveiled the collaboration at Singapore International Cyber Week, with analysis running entirely on the user's device.
How does the Gemma-based scam detection model protect user privacy?
The model processes conversation data on the handset rather than sending it to a cloud server, so it can operate inside end-to-end encrypted messaging apps. Training excluded private or personally identifiable citizen data, using anonymised material from scam reports and intelligence feeds inside a GovTech-managed Google Cloud environment.
Why is Singapore investing in conversational scam detection?
Scams make up over half of Singapore's reported crimes and have cost more than S$4 billion since 2019, according to the announcement. Social-engineering schemes, where attackers build credible conversations before requesting money, now account for up to 80% of scam losses, and generic models often miss localised deception techniques.
When will the on-device scam detection tool launch in Singapore?
No public release date has been disclosed. Minister Josephine Teo did not provide a launch timeline or performance figures such as detection and false-positive rates, and details of integration into Android or specific messaging platforms remain undisclosed, according to CXO Digitalpulse.
How does the two-stage pipeline save smartphone battery life?
A lightweight, low-power classifier first screens incoming messages for potential threats using minimal energy. Only conversations flagged as suspicious move to the fine-tuned Gemma model, which analyses a broader message sequence to confirm manipulation or dismiss a false alarm, balancing detection capability with smartphone processing and battery constraints.