GovTech Singapore and Google Deploy On-Device Gemma to Detect Conversation-Based Scams

Singapore's Government Technology Agency and Google DeepMind announced a fine-tuned Gemma model for Android phones on Wednesday that analyzes conversation content entirely on the device, with no data sent to external servers. The system, revealed by Minister for Digital Development and Information Josephine Teo at Singapore International Cyber Week 2026, targets a category of fraud that has proven structurally immune to existing defenses: social-engineering scams built through extended private conversations rather than malicious links or fake websites.

Singapore's scam problem is significant by any measure. The country has recorded more than S$4 billion in scam-related losses since 2019, and scams now account for more than half of all reported crimes. What makes the figure especially significant is that roughly 80 percent of those losses are attributed to social-engineering fraud — the kind in which a scammer impersonates a bank representative, government official, or local authority figure over multiple messages, builds a plausible narrative, manufactures urgency, and eventually persuades a victim to transfer money or share credentials. These scams contain no malicious URLs, no unusual file attachments, and no detectable code signatures. They are, at the technical level, indistinguishable from a normal conversation — which is exactly why they have been untouchable by every anti-scam system deployed before this one.

Why SMS Filters and URL Blockers Cannot Stop Conversational Fraud

Singapore already operates a layered anti-scam infrastructure. SATIS, the Scam Analytics and Tactical Intervention System built by GovTech and the Singapore Police Force, scans more than 600,000 websites daily and takes down roughly 20,000 malicious pages monthly. ScamShield, the consumer-facing app built by the government's Open Government Products team, filters suspicious SMS and incoming calls at the network level. A separate rMSE system scans web properties for emerging threats. GovTech is also the first government agency worldwide to join the Global Signal Exchange — a real-time threat intelligence platform co-founded by Google, Oxford Information Labs Research, and the Global Anti-Scam Alliance, which tracks more than 400 million threat signals and counts Meta and Microsoft as members.

What none of these systems can do is read a WhatsApp or Telegram conversation. Network-level SMS filtering sees message metadata and, in some cases, message bodies for unencrypted texts, but the conversational scam category has largely migrated to end-to-end encrypted messaging applications, where no external system — not the carrier, not a government server, not a cloud API — can inspect the content. The only computational context that can examine an encrypted conversation is the device itself. That architectural reality is what the GovTech-Google DeepMind project is built around.

A Two-Stage Pipeline Built Around Battery and Privacy Constraints

The detection system uses a two-stage on-device pipeline specifically designed to address the competing constraints of mobile hardware: the need to screen large message volumes cheaply while reserving computationally expensive LLM inference for cases that warrant it.

The first stage is a lightweight classifier that monitors incoming messages continuously. It is optimized for power efficiency and low latency, functioning as a triage layer rather than a detection system in its own right. When the classifier identifies a conversation that matches patterns associated with suspicious engagement — escalating requests for personal information, pressure tactics, impersonation cues — it forwards the relevant conversation snapshot to the second stage.

The second stage is a fine-tuned Gemma model, Google DeepMind's open-weight model family designed in part for edge and mobile deployment. Where the lightweight classifier looks for individual signal features, the Gemma model analyzes the full arc of a conversation: the sequence of messages, the roles each party plays, the escalation of requests over time, and the rhetorical tactics used to suppress the recipient's skepticism. When the model identifies a manipulation pattern with sufficient confidence, it surfaces a real-time warning to the user — a prompt to pause and reconsider before sharing money or credentials.

The pipeline design matters because it directly answers the battery-drain objection that has historically limited on-device AI security tools. Running a full LLM inference pass on every incoming message would be prohibitively expensive on a smartphone. By inserting a cheap gating stage first, the system ensures the Gemma model fires only on a small fraction of messages: those that have already passed a lower-cost suspicion threshold.

Training Inside Singapore's Sovereign Sandbox

The model was fine-tuned inside what GovTech describes as a Sovereign Sandbox — a Google Cloud environment provisioned and managed by GovTech Singapore, within which training data from Singapore-specific scam corpora remains under Singaporean operational control. Google provides the Gemma base weights and the cloud infrastructure; GovTech retains custody of the training data and the resulting fine-tuned model. Full governance details are described in Google DeepMind's announcement.

The design reflects a tension that has complicated AI adoption by governments across Southeast Asia: frontier model capabilities are concentrated in a small number of American and Chinese technology companies, but the data needed to make those models useful for domestic law enforcement and public safety applications is politically and legally sensitive. The Sovereign Sandbox model threads that needle by allowing GovTech to supply locally collected, anonymized scam data — drawn from police reports and intelligence feeds, with all personally identifiable information excluded — without those datasets entering Google's general training infrastructure.

Mark Chen, Director of Trust and Safety at GovTech Singapore, framed the design philosophy directly: "Fighting scams shouldn't mean sacrificing personal privacy." The statement captures both a product constraint and a political one. Singapore has faced public debate about the expansion of state surveillance tools, and any anti-scam system that required routing private messages through government servers would have encountered significant resistance. The on-device approach sidesteps that objection: neither Google nor GovTech can access the conversations the model analyzes, because those conversations never leave the device.

How This Fits Into Google's Broader On-Device Scam Detection Push

The GovTech collaboration is not Google DeepMind's first use of Gemma or Gemini Nano for consumer scam protection, but it represents a significant expansion in scope and partnership depth. Google's prior on-device scam detection work has been focused on specific, narrow categories: Gemini Nano in Google Messages catching scam SMS patterns; Gemini Nano in Chrome 137 flagging tech-support scam pages by detecting signatures like the Keyboard Lock API; call-screening features on Pixel devices that analyze audio for scam-call indicators.

Each of those systems targets a discrete threat surface: a specific message type, a specific browser context, a specific call scenario. What the GovTech partnership introduces is a system trained to recognize manipulation as a conversational strategy rather than a technical artifact — to detect the social and rhetorical structure of a scam, not just its technical fingerprints.

That distinction matters technically because social-engineering scams specifically avoid technical fingerprints. A scammer impersonating a Singapore Police Force officer over WhatsApp does not send a phishing link; they send a series of ordinary-looking text messages that over several exchanges establish credibility, create urgency, and guide the victim toward voluntary disclosure. No URL database or malware signature catches that sequence. A language model trained on local scam corpora — impersonations of DBS Bank, CPF Board, Singapore Police Force — has a reasonable chance of recognizing the pattern before the victim reaches the extraction point.

Limitations and What Remains Unverified

Several details about the deployment remain undisclosed or could not be independently verified. The specific Gemma variant used — the model's parameter count, quantization depth, and inference latency on typical hardware — has not been published. No accuracy metrics were announced; the detection rate and false-positive rate under real-world conditions are unknown.

The rollout timeline is also unspecified. Multiple sources indicate the project has no confirmed public availability date, and that reaching users at scale will require integration agreements with messaging platforms — a dependency that makes the deployment timeline difficult to predict. The system has been demonstrated and announced but has not yet been deployed to end users.

Whether the detection generalizes beyond the training distribution — to new scam variants, new impersonation targets, or scam conversations conducted in code-switching languages like Singlish — is a genuine open question that will require real-world evaluation to answer. Language capabilities of the fine-tuned model were not disclosed in the announcement.

What Singapore's Model Could Mean for the Region

The Sovereign Sandbox approach, if it demonstrates measurable results in Singapore's deployment, offers a blueprint for other Southeast Asian governments facing similar constraints. Countries including Malaysia, Thailand, and Vietnam face comparable social-engineering scam epidemics, similarly high smartphone penetration rates, and similar political obstacles to routing citizen communications through cloud-based government surveillance infrastructure. Singapore's relatively small population and advanced digital infrastructure make it a practical testbed for approaches that other nations in the region might adopt.

Read more: Singapore opens the world's first humanoid robotics centre for public safety AI

The partnership also represents a data point in the broader competition between frontier AI companies for government relationships. Google's National AI Partnership with the Singapore Government provides a reference deployment that no competitor currently has in the consumer scam-detection space. For both GovTech and Google DeepMind, the system's real-world performance — once it reaches users — will determine whether the Sovereign Sandbox model evolves from a national experiment into a regional standard.