# GovTech and Google DeepMind Unveil On-Device Scam Detection Model at SICW 2026 Source: TechNewsList (https://technewslist.com) Canonical URL: https://technewslist.com/en/article/govtech-deepmind-on-device-scam-model-sicw-2026-10-14-morning Section: AI (https://technewslist.com/en/ai) Author: TechNewsList Language: en Published: 2026-10-14T05:38:30.711+00:00 Updated: 2026-10-14T05:38:30.888925+00:00 > Singapore and Google DeepMind have introduced an on-device scam detection model fine-tuned from Gemma, targeting social engineering fraud in encrypted chats without server data leaks. ## TL;DR - GovTech Singapore and Google DeepMind unveiled an on-device AI model fine-tuned from Gemma to detect social engineering scams in real time. - The defense operates entirely on client devices with zero cloud telemetry egress, preserving the privacy of end-to-end encrypted chats. - Singapore reported over S$4 billion in fraud losses since 2019, with social engineering accounting for eighty percent of damages. ## Key points - The project was unveiled on October 14 at SICW 2026 by Minister for Digital Development and Information Josephine Teo. - Google DeepMind compressed and fine-tuned open Gemma weights using localized Singaporean scam transcripts and impersonation patterns. - On-device inference inspects conversational manipulation trajectories locally rather than scanning message content on remote cloud servers. - The client-side safety guard triggers non-intrusive warnings directly before banking transactions or credential handoffs occur. - Singapore concurrently expanded its national defensive perimeter by joining Google's Fairwind Programme for advanced cybersecurity testing. ## What happened At the eleventh annual Singapore International Cyber Week (SICW 2026) in Singapore, Minister for Digital Development and Information Josephine Teo announced an official collaboration between the Government Technology Agency of Singapore (GovTech) and Google DeepMind to deploy a localized, on-device artificial intelligence scam-detection model. The system is designed to evaluate real-time messaging conversational dynamics directly on user smartphones, detecting subtle psychological manipulation tactics used by professional fraud syndicates. According to coverage by [Channel NewsAsia](https://www.channelnewsasia.com/singapore/govtech-google-deepmind-ai-scam-quantum-sicw-6455621), the initiative leverages Google DeepMind's open-weight Gemma model family. The model has undergone specialized fine-tuning on regional fraud patterns, fake governmental notices, and impersonation scripts frequently encountered across Southeast Asia. The engineering team configured the weights specifically for local Android device runtimes, allowing behavioral evaluation to execute without sending raw message transcripts to remote infrastructure. ![On-device Gemma scam detection pipeline](https://rkhynbcsbnkkcwgexzwg.supabase.co/storage/v1/object/public/media/api/1791956303551-1a7oim-govtech-deepmind-on-device-scam-model-sicw-2026-10-14-morning-inside-1-9e650f2aee.webp) ## Why it matters Digital fraud has escalated into a systemic economic threat across urban hubs. In Singapore alone, authorities report that scams have generated more than four billion Singapore dollars in aggregate consumer losses since 2019. Traditional perimeter defenses, such as domain filtering and telecom-level SMS blacklists, increasingly struggle against contemporary criminal syndicates that operate across peer-to-peer encrypted messaging platforms like WhatsApp, Telegram, and Signal. Because modern chat ecosystems enforce end-to-end encryption, network gateways and telecommunications carriers cannot inspect message payloads in transit without compromising the cryptographic privacy of honest conversations. As detailed by [Unite.AI](https://www.unite.ai/google-deepmind-and-govtech-unveil-on-device-scam-detection-model/), moving the intelligence runtime directly into the client operating system circumvents this impasse. Users receive the protective benefits of automated conversational auditing while maintaining absolute mathematical confidentiality against third-party intermediaries and cloud operators. Furthermore, the announcement coincides with the formal publication of Singapore's updated national cyber policy, detailed by the [Cyber Security Agency of Singapore](https://www.csa.gov.sg/news-events/press-releases/singapore-launches-cybersecurity-strategy-2026-to-strengthen-national-cyber-resilience/). The strategy establishes active, automated machine learning defenses as a mandatory tier of public cyber resilience alongside infrastructure hardening. ## Technical details The technological engine of the GovTech and DeepMind partnership centers on quantized instances of the lightweight Gemma open model architecture. Rather than executing generalized natural language generation, the model has been pruned and distilled to function as an intent and manipulation classifier that continuously parses rolling contextual dialogue buffers within active message threads. Training datasets were constructed from verified forensic archives compiled by Singaporean law enforcement agencies. These corpuses incorporate hundreds of thousands of conversational exchanges documenting romance scams, fake investment schemes, public agency impersonations, and credential harvesting rituals. By mapping the linguistic stages through which attackers build trust, induce panic, and isolate victims, the classifier identifies escalating pressure gradients. To operate within typical smartphone power and thermal envelopes, the model utilizes specialized integer quantization (INT4 and INT8) targeting Qualcomm Snapdragon and MediaTek Neural Processing Units (NPUs). Inference latency remains restricted to millisecond intervals, executing opportunistically upon incoming message rendering without causing user interface jitter or noticeable battery exhaustion. ![SICW 2026 Cyber Exhibition](https://rkhynbcsbnkkcwgexzwg.supabase.co/storage/v1/object/public/media/api/1791956305840-1932yv-govtech-deepmind-on-device-scam-model-sicw-2026-10-14-morning-inside-2-866d902207.webp) ## Market / industry impact The joint venture signals a fundamental architectural migration in enterprise and consumer threat intelligence toward decentralized client inference. For years, major cybersecurity software vendors required telemetry pipelines that harvested customer emails and chat snippets into central threat analysis lakes. This legacy approach faces severe friction under modern privacy regulations, including Europe's General Data Protection Regulation and cross-border data transfer statutes. By establishing that a compact on-device language model can identify conversational deception with high precision, Google DeepMind and GovTech provide a viable template for consumer tech manufacturers. Mobile operating system developers and hardware vendors are under intensifying consumer demand to integrate native, privacy-preserving defenses directly into hardware silicon. Concurrently, the Singaporean government confirmed that its national cybersecurity operational units are participating in Google's Fairwind Programme. This arrangement provides accredited government defenders with early sandboxed access to cutting-edge frontier defensive models, creating a coordinated bridge between national intelligence operations and client consumer protection. ## What to watch next While the technology prototype has been demonstrated to delegates at SICW 2026, several operational milestones remain before public deployment across commercial device fleets. Engineers must validate the false-positive intervention rate in multi-lingual environments where colloquial phrasing, such as Singaporean English (Singlish), frequently combines multiple dialects in informal syntax. Regulatory authorities and independent privacy auditing bodies will also inspect the on-device software package to verify that no latent background telemetry leaks conversational tokens back to corporate cloud partitions. The public beta evaluation is anticipated to commence across municipal testing cohorts in early 2027. Longer term, security analysts will monitor whether criminal syndicates adapt by deploying counter-evasion prompts designed to confuse on-device classifiers. The battle against social engineering fraud is rapidly pivoting into an edge-computing adversarial competition where defensive neural models on handheld devices face adversarial prompt injections from criminal operations. ## Sources - [Cyber Security Agency of Singapore](https://www.csa.gov.sg/news-events/press-releases/singapore-launches-cybersecurity-strategy-2026-to-strengthen-national-cyber-resilience/) — Official government cybersecurity policy and defensive machine learning guidelines. - [Channel NewsAsia](https://www.channelnewsasia.com/singapore/govtech-google-deepmind-ai-scam-quantum-sicw-6455621) — Coverage of Minister Josephine Teo's keynote speech and the on-device Gemma architecture. - [Unite.AI](https://www.unite.ai/google-deepmind-and-govtech-unveil-on-device-scam-detection-model/) — Architectural deep dive into on-device Gemma fine-tuning, hardware NPU requirements, and zero-cloud telemetry. Mentions: GovTech Singapore, Google DeepMind, Josephine Teo, Cyber Security Agency of Singapore ## Sources - [Cyber Security Agency of Singapore](https://www.csa.gov.sg/news-events/press-releases/singapore-launches-cybersecurity-strategy-2026-to-strengthen-national-cyber-resilience/) - [Channel NewsAsia](https://www.channelnewsasia.com/singapore/govtech-google-deepmind-ai-scam-quantum-sicw-6455621) - [Unite.AI](https://www.unite.ai/google-deepmind-and-govtech-unveil-on-device-scam-detection-model/)