Singapore and Google DeepMind Deploy On

Google DeepMind and Singapore's Government Technology Agency (GovTech Singapore) have unveiled a privacy-preserving scam detection system that runs entirely on a user's smartphone — analyzing suspicious messages in real time using Google's open Gemma model , without transmitting any conversation data to a server.

The collaboration, announced at Singapore International Cyber Week (SICW 2026), which runs through October 15 at Sands Expo and Convention Centre under the theme "Cyber at a Crossroads: Shaping the New Cyber Order," represents the first major technical deliverable of the National AI Partnership between Google DeepMind and the Singapore government. That bilateral program, announced in May 2026 under Singapore's Ministry of Digital Development and Information (MDDI), now has its first concrete engineering output.

Singapore Lost $321 Million to Scams in the First Half of 2026 Alone

The problem the system is designed to address is not hypothetical. Singapore lost S$410.6 million (approximately $321 million USD) to scam activity in the first half of 2026 — a 17.9 percent year-on-year decline that reflects intensifying government countermeasures but remains deeply significant at scale. Scams now account for more than half of all reported crimes in Singapore, and cumulative losses since 2019 have exceeded S$4 billion (approximately $3.13 billion USD).

The financial damage is concentrated in two categories. Investment fraud claimed S$169.8 million (approximately $133 million USD) in the first half of 2026; government official impersonation scams contributed another S$90.8 million (approximately $71 million USD). What makes these schemes resistant to conventional defenses is their method: social engineering — sustained, multi-turn conversation designed to psychologically manipulate a victim into voluntarily transferring money or credentials — is estimated to drive approximately 80 percent of total financial losses. An estimated 80.8 percent of victims transferred funds willingly, having been deceived through psychological manipulation rather than directly defrauded through account compromise.

This conversational manipulation pattern is precisely what rule-based keyword filters and blocklists cannot reliably catch. Detecting it requires reasoning over the arc of a conversation — not just the content of any single message — which demands a capable language model, not pattern-matching logic.

How a Two-Stage Pipeline Catches Scams Without Draining Your Battery

The core engineering challenge was reconciling two competing constraints: nuanced conversational reasoning, which requires a capable language model, and continuous on-device operation , which cannot deplete a smartphone battery.

The solution is a two-stage inference pipeline that separates the cheap classification problem from the expensive reasoning problem.

Stage 1 deploys a lightweight, power-efficient classifier that screens every incoming message in real time. This first-pass model handles the bulk of traffic — returning a benign verdict on ordinary messages without engaging the heavier compute layer, keeping background power draw low. Stage 2 activates only when Stage 1 flags a potentially suspicious message. At that point, a fine-tuned Gemma model ingests a larger snapshot of the full conversation, including prior turns, to conduct in-context multi-turn analysis. It determines whether the conversational arc reflects genuine social engineering or a false alarm, and when a scam is detected, the system surfaces a real-time warning on the device — advising the user to pause before sharing sensitive information or initiating a funds transfer.

The choice of Gemma as the Stage 2 model reflects specific on-device engineering requirements. Gemma is Google DeepMind's family of open, lightweight models purpose-built for edge and mobile deployment. The Gemma 4 generation, released in April 2026 under an Apache 2.0 open license, includes edge-optimized variants designed for mobile-class hardware — with approximately 2 gigabytes of on-device storage required. Larger server-resident models cannot meet the power and latency constraints of continuous on-device inference; Gemma's architecture is purpose-built for exactly this class of deployment.

What Privacy Guarantee Does This System Actually Provide?

The privacy architecture is absolute by design: no message content leaves the device at any point during analysis. The model accesses users' private communications only with the user's explicit consent, in accordance with the application's terms of use, and neither Google nor the Singapore government can access users' private communications through the system.

This represents a structural departure from conventional threat-detection approaches, which typically rely on server-side message scanning — a method that creates both privacy exposure and latency, and that conflicts directly with end-to-end encrypted messaging platforms. The on-device model resolves that conflict by keeping analysis local. A messaging platform can deploy this system without modifying its encryption architecture or routing messages through external servers.

The training pipeline was built under equally stringent data governance. Fine-tuning took place inside a Sovereign Sandbox — GovTech Singapore's own managed and maintained Google Cloud Platform environment, keeping all training data isolated from external systems. No personal data or personally identifiable information was used at any stage. Training datasets were derived exclusively from anonymized scam reports and intelligence feeds, and the model was stress-tested against synthetic dialogues and locally curated corpora covering a wide range of scam categories and impersonation tactics specific to Singapore's context.

Mark Chen, Director of Trust and Safety at GovTech Singapore, framed the dual imperative directly: "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."

Why Local Fine-Tuning Catches What Generic Scam Classifiers Miss

The decision to fine-tune Gemma on Singapore-specific data, rather than deploy a globally trained scam classifier, is the system's most technically significant design choice.

Generic models struggle with local deception techniques because their training corpora may not adequately represent the specific institutions being impersonated, the vernacular in use, or the regional schemes in circulation. A model capable of detecting standard phishing patterns may entirely miss an impersonation of a Singapore public housing authority official or a locally prominent bank — because those specific social and institutional contexts were underrepresented in its original training data.

The fine-tuned Gemma model was trained on local scam conversations spanning impersonations of Singaporean public institutions, banks, and prominent local figures, as well as regionally trending fraud schemes. This localization allows the model to parse the difference between ordinary social conversation and genuine manipulation in the specific cultural and linguistic context where it operates — a distinction that a globally trained model would systematically miss.

The joint blog post, authored by DeepMind engineers JJ Geewax and Miguel de Andrés, notes that the collaboration intends to more details in coming months and explicitly flags the potential for similar technology to be deployed in other countries. That framing suggests the Sovereign Sandbox model — privacy-respecting, nationally specific fine-tuning using sovereign cloud infrastructure — may be designed for replication as much as for Singapore specifically. It represents a template for how national governments can train locally calibrated AI models under data-sovereignty constraints without sharing citizen data with a commercial AI partner.

Deployment Is Phased; No Consumer Release Date Has Been Set

The rollout of the model into consumer applications will proceed in phases, developed in ongoing partnership with the Singapore government. No timeline for broad availability was announced at SICW 2026. The phased approach reflects the sensitivity of deploying real-time conversational classifiers within messaging ecosystems, where false positive rates — incorrectly flagging legitimate conversations as scams — can erode user trust and create friction for everyday communications.

The governance framework governing the collaboration also provides a potential model for messaging platforms navigating the tension between user protection and end-to-end encryption. By keeping analysis on-device, the system gives platforms a path to deploying safety features that would otherwise conflict with their encryption architectures. Platforms operating in Singapore face a concrete incentive: proposed amendments to the Online Criminal Harms Act (OCHA), introduced in Parliament in August 2026, would raise maximum fines for non-compliant platforms from S$1 million to S$10 million (approximately $7.8 million USD). The amendment had not been enacted as of the date of this article.

Singapore's AI Minister Has Made Scam-Resistant Infrastructure Central to a National Policy Agenda

The SICW 2026 deployment is the first concrete technical output of a national AI strategy whose scope extends well beyond scam detection.

Josephine Teo, Singapore's Minister for Digital Development and Information, has positioned Singapore as a global reference point for responsible AI governance across a range of technology policy domains. In January 2026, Singapore became the first government anywhere to issue a Model Governance Framework specifically for agentic AI systems — the class of AI capable of taking autonomous actions, not merely generating responses. In March 2026, Teo launched the National AI Impact Programme , a structured initiative targeting 10,000 Singapore enterprises and the upskilling of 100,000 workers in AI-bilingual competencies. Singapore has also imposed mandatory sustainability reporting requirements on data centers operating in the country, and the city-state hosted the second session of the United Nations Open-Ended Working Group on ICT security, where Teo has been a prominent voice on international AI governance norms.

Google DeepMind's national partnership strategy reflects a parallel trajectory. The company announced AI partnerships with India in February 2026 and South Korea in April 2026 before announcing the Singapore agreement in May. The three agreements together represent a systematic effort to build government-partnered AI infrastructure across Asia's largest and most digitally developed economies.

Singapore's own enforcement posture adds further context. In the first half of 2026, Singapore Police Force disrupted more than 47,500 mobile lines, 37,500 WhatsApp lines, 31,600 online accounts, and 52,200 scam-related websites. That level of operational investment — combined with the legislative pressure of the proposed OCHA amendments and the Ministry's AI policy agenda — makes Singapore a natural testing ground for AI-augmented consumer protection at national scale.

The Sovereign Sandbox Model Is the Architecture Other Governments Are Watching

Two converging trends in applied machine learning meet in this deployment.

The first is the maturation of on-device inference for tasks previously assumed to require server-side compute. The two-stage pipeline — a cheap classifier as a pre-filter, with a capable language model reserved for ambiguous edge cases — is an engineering pattern being adopted increasingly across mobile AI workloads as lightweight model families like Gemma become capable of nuanced reasoning on consumer hardware. The pattern resolves the fundamental tension between model capability and power consumption in a way that earlier single-model approaches could not.

The second, and more structurally significant, is the emergence of the Sovereign Sandbox as a model of public digital infrastructure. Rather than relying on general-purpose models trained on global internet data, national governments are beginning to build — with commercial AI labs as technical partners — locally calibrated models fine-tuned on nationally specific threat data, operated under data-sovereignty constraints, and deployed for directly civic ends. GovTech Singapore's managed cloud environment represents one practical implementation of that approach: a government retains full control over its citizen data while accessing the technical infrastructure and model capacity of a global AI lab.

Whether the same template will prove adaptable to countries with different regulatory environments, different languages, and different encryption-policy landscapes remains to be demonstrated. But Singapore's willingness to move from the May 2026 partnership announcement to a working deployment unveiled on an international cybersecurity stage in roughly five months signals that the government views the model as exportable — and has deliberately chosen one of Asia-Pacific's most visible cybersecurity venues to make that case.

Currency conversions throughout this article are approximate; rates referenced as of October 14, 2026.

Frequently Asked Questions

What is the Sovereign Sandbox, and why does it matter beyond Singapore?

The Sovereign Sandbox is GovTech Singapore's managed Google Cloud Platform environment used to fine-tune the Gemma model on local scam data, with training kept entirely isolated from external systems and no personal or personally identifiable information used. Its significance is architectural: it establishes a replicable template for how national governments can train locally calibrated AI models under data-sovereignty constraints — using a commercial AI lab's infrastructure and expertise — without sharing citizen data with that partner.

The joint blog post from Google DeepMind engineers JJ Geewax and Miguel de Andrés explicitly notes plans to more details in coming months and flags the potential for similar deployments in other countries, signaling that the Sovereign Sandbox approach is being positioned as an exportable model of national AI governance, not only a Singapore-specific solution.

How does this differ from Singapore's existing ScamShield app?

Singapore's existing ScamShield application cross-references incoming phone numbers and links against a government-maintained database of known scam identifiers — a rule-based approach that can only intercept previously documented threats. The new system uses a fine-tuned language model to analyze the full conversational arc of a message exchange, detecting novel social engineering patterns even when the number or link has never previously been flagged. In principle, it can identify a first-time scammer using a legitimate phone number, provided that person's conversational behavior matches the manipulation patterns in training data.

Does on-device analysis work on end-to-end encrypted messaging platforms?

Yes — by design. Because all analysis happens on the device, the system does not require access to a message stream at the server level and does not require breaking end-to-end encryption. A messaging platform can integrate the on-device model without modifying its encryption architecture or routing messages through an external server. This makes the approach compatible with platforms that currently cannot deploy server-side scanning without violating their own privacy guarantees — and potentially relevant to the regulatory pressure platforms face under Singapore's proposed OCHA amendments.

When will this be available to people in Singapore?

No specific consumer release date was announced at SICW 2026. The rollout will be phased, developed in ongoing partnership with the Singapore government, with particular attention to keeping false positive rates low — incorrect scam alerts on legitimate conversations could erode the trust the system depends on. The collaboration has indicated it will further details in the coming months.