Google DeepMind and Singapore GovTech develop on-device scam detection AI using Gemma
Google DeepMind and Singapore's Government Technology Agency (GovTech) are developing an AI-based approach to scam detection that uses Google's Gemma models to identify potentially fraudulent messages directly on mobile devices. The initiative focuses on combining language-model analysis with on-device processing, allowing suspicious content to be evaluated without routinely transmitting the underlying conversation to an external AI server.
The approach addresses a growing problem for conventional anti-scam systems: fraudulent messages increasingly resemble ordinary communications, often avoiding the suspicious links, repetitive wording and recognizable templates used by earlier phishing campaigns.
An on-device language model can examine the meaning of a message, identify requests for sensitive information and recognize manipulation patterns that simple keyword filters may miss. However, the technology also faces important constraints involving model accuracy, mobile hardware, privacy permissions and the risk of incorrectly flagging legitimate communications.
The collaboration is a development initiative, not confirmation that a new scam-detection service has already been released to every smartphone user. Publicly established information does not yet provide a universal rollout date, a complete list of supported devices or independently verified detection rates.
··········
SINGAPORE'S SCAM LOSSES EXCEED S$4 BILLION, WITH SOCIAL ENGINEERING DRIVING MOST OF THE DAMAGE.
The project was announced on October 14, 2026, during Singapore International Cyber Week, as part of the National AI Partnership between Google DeepMind and the Singapore Government.
According to Google DeepMind, scams account for more than half of Singapore's reported crimes and have generated over S$4 billion in cumulative losses since 2019. Social-engineering schemes account for as much as 80% of scam-related financial losses.
These attacks often unfold through conversations rather than isolated messages. A fraudster may impersonate a bank employee, public official or trusted individual, establish credibility over several exchanges and only later request a transfer or confidential information.
The delayed request creates a detection challenge. An individual message may appear harmless even when the wider conversation reveals a pattern of manipulation.
The new system is designed to identify that broader pattern before a victim shares credentials or transfers funds.
........
Characteristic | Announced information |
Developers | Google DeepMind and GovTech Singapore |
Announcement date | October 14, 2026 |
AI model family | Gemma |
Primary task | Conversational scam detection |
Processing location | User's smartphone |
Detection architecture | Two-stage classification |
Training data | Anonymized local scam reports and intelligence |
Public accuracy metrics | Not disclosed |
Consumer availability | No confirmed general rollout |
Model parameter count | Not disclosed |
........
The figures describing Singapore's fraud problem establish the scale of the opportunity, but they do not measure the effectiveness of the new detector. A reduction in financial losses would need to be demonstrated through deployment data rather than inferred from the existence of the model.
··········
A TWO-STAGE PIPELINE COMBINES FAST MESSAGE SCREENING WITH GEMMA-BASED REASONING.
The system uses a lightweight classifier for initial screening and activates a fine-tuned Gemma model when more detailed conversational analysis is necessary.
This design addresses a central limitation of running language models continuously on smartphones: battery consumption and computational overhead.
Rather than sending every incoming message through the more demanding model, the first stage performs a relatively inexpensive assessment.
Messages classified as benign can pass without further analysis. Potentially suspicious messages trigger the second stage, which examines a larger portion of the conversation.
The processing sequence is: incoming message, local lightweight classifier, and, only for potentially suspicious content, multi-turn analysis by Gemma followed by a scam warning or benign classification. Both stages are intended to operate on the device.
The second stage evaluates conversational context rather than treating each message independently.
For example, a message asking someone to confirm their identity may be legitimate when sent through an established support process. The same request may be suspicious when preceded by unsolicited contact, pressure to act immediately and instructions to use an unfamiliar payment channel.
Gemma's role is to help distinguish these situations by examining relationships between messages.
The two-stage approach also creates an important engineering trade-off.
A highly selective first-stage classifier can reduce unnecessary Gemma invocations and save battery power, but it may overlook sophisticated attacks. A more sensitive classifier may catch additional threats while requiring the second stage to process more ordinary conversations.
The optimal balance depends on detection performance, latency, device energy consumption and the consequences of missed scams.
Google has not disclosed the precise first-stage model, its screening thresholds or the percentage of messages expected to reach Gemma.
··········
GEMMA IS FINE-TUNED ON LOCAL SCAM CONVERSATIONS RATHER THAN GENERIC PHISHING EXAMPLES.
Google DeepMind adapted a lightweight Gemma model using scam examples reflecting Singapore's local fraud environment.
The training material includes impersonation attempts involving public institutions, banks and recognizable local figures, as well as region-specific fraud schemes.
This localization is important because fraud patterns depend on language, cultural expectations, institutional relationships and the services commonly used by victims.
A generic scam classifier may recognize a suspicious request for a password while failing to understand why a particular government impersonation or payment instruction is deceptive in a specific country.
Fine-tuning provides a mechanism for adapting a pretrained language model to these local patterns without training an entirely new foundation model.
The objective is to improve the model's ability to distinguish normal social interactions from conversations designed to manipulate the recipient.
Google has not identified the exact Gemma checkpoint, its parameter count, quantization format or the number of training examples.
Consequently, it is not possible to establish whether the system uses a particular published Gemma variant or to calculate its expected memory footprint from the announcement.
The company also has not released the fine-tuned checkpoint for independent testing.
These omissions do not invalidate the architecture, but they limit the extent to which outside researchers can reproduce its performance.
··········
ON-DEVICE PROCESSING AIMS TO PRESERVE PRIVATE CONVERSATIONS AND END-TO-END ENCRYPTION.
The privacy objective is to analyze messages on the user's device without routinely uploading their contents to a cloud-based detection service.
Many messaging applications use end-to-end encryption, meaning that message contents are intended to be accessible to the communicating parties rather than the service provider.
Cloud-based scanning can create tension with that design if it requires private messages to be transmitted to another service for inspection.
On-device inference offers a different approach.
Once a message is available to the authorized application on the recipient's device, a local model can evaluate it without requiring the provider to decrypt conversations on its servers.
This can preserve the confidentiality benefits of end-to-end encryption while enabling additional protection at the endpoint.
However, local processing does not eliminate every privacy concern.
The application still needs appropriate access to conversation content. Its permissions, storage behavior, telemetry and update mechanisms determine how private information is handled.
The distinction between model inference and application-level data processing is therefore important.
A model can execute entirely offline while the surrounding application still communicates with remote services for unrelated purposes.
Google and GovTech describe local conversation processing as a central design property, but the complete privacy behavior of a future consumer implementation will depend on the application in which the technology is integrated.
The announcement does not establish that the system can automatically inspect every encrypted messaging application. Operating-system restrictions and application permissions will determine which conversations can be analyzed.
··········
TRAINING TOOK PLACE IN A GOVTECH-MANAGED SOVEREIGN CLOUD ENVIRONMENT.
Google DeepMind describes three governance measures adopted during development.
First, model fine-tuning took place within a Google Cloud environment managed and maintained by GovTech Singapore.
Second, the training process excluded private or personally identifiable citizen information, using anonymized material derived from scam reports and intelligence feeds.
Third, the system was stress-tested against synthetic dialogues and local datasets representing different fraud categories and evolving impersonation techniques.
........
Governance measure | Announced implementation | Practical objective |
Training environment | GovTech-managed Google Cloud | Controlled model development |
Training information | Anonymized scam reports and intelligence | Reduce exposure of personal information |
Evaluation material | Synthetic dialogues and local corpora | Test diverse fraud scenarios |
User-side inference | On-device processing | Avoid routine cloud transmission of chats |
Response to suspicious activity | Real-time warning | Encourage intervention before loss |
........
The separation between training and inference is important.
Using a managed cloud environment to fine-tune the model does not contradict the objective of executing the finished detector locally.
Training generally requires more computational resources than running a compact, optimized model on a smartphone.
Nevertheless, anonymization quality, dataset governance and access controls remain important when working with real fraud reports.
The announcement describes the measures used but does not provide an independently audited account of the training dataset or complete evaluation procedures.
··········
THE MODEL WARNS USERS BEFORE SENSITIVE INFORMATION OR MONEY IS SHARED.
When the detector identifies a potentially manipulative conversation, the intended response is a real-time warning displayed on the user's device.
The warning advises the recipient to reconsider the interaction before transferring funds or disclosing sensitive information.
This intervention is different from automatically blocking a message or reversing a payment.
The system is designed to support the user's decision at a potentially consequential moment.
For example, a fraudster may initially claim to represent a bank and gradually introduce an urgent request to move funds into a supposedly protected account.
An effective detector would need to recognize the developing manipulation before the transfer occurs.
Timing matters because a warning delivered after the victim has sent money may have limited preventive value.
The wording of the warning also matters.
An excessively alarming message may confuse users, while a vague notification may be ignored. Repeated false warnings can reduce trust in the system.
Google has not disclosed the precise warning interface, whether users can submit feedback or how applications will handle repeated alerts.
These design choices will influence real-world effectiveness alongside the underlying model's classification performance.
··········
FALSE POSITIVES AND MISSED SCAMS WILL DETERMINE REAL-WORLD EFFECTIVENESS.
Google DeepMind has not published numerical results for scam-detection accuracy, precision, recall or false-positive rates for this project.
The absence of these measurements prevents a quantitative comparison with established anti-fraud systems.
A scam detector must balance two types of error.
A false negative occurs when a fraudulent conversation is classified as harmless. A false positive occurs when a legitimate conversation is incorrectly identified as suspicious.
Both have costs.
False negatives can expose users to financial losses or credential theft. False positives can interrupt legitimate conversations and undermine confidence in the warning system.
The consequences also depend on how frequently users encounter genuine scams.
Consider a hypothetical system with 95% sensitivity and 98% specificity, operating in an environment where 1% of evaluated conversations are scams.
Across 10,000 conversations, 100 would be fraudulent and 9,900 legitimate.
The model would detect approximately 95 scams but also flag about 198 legitimate conversations.
........
Hypothetical outcome | Conversations |
Total evaluated | 10,000 |
Actual scams | 100 |
Correctly detected scams | 95 |
Missed scams | 5 |
Legitimate conversations | 9,900 |
Incorrectly flagged legitimate conversations | 198 |
Precision of scam alerts | Approximately 32.4% |
........
This is a mathematical illustration, not a measurement of Google's detector.
It demonstrates why seemingly high sensitivity and specificity do not necessarily produce highly reliable alerts when genuine scams are relatively uncommon.
In the example, fewer than one-third of warnings would correspond to actual scams.
A deployed system therefore needs evaluation against realistic fraud prevalence, including benign conversations that resemble scam patterns.
Testing should also examine different languages, demographics, communication styles and attack types.
Scammers may adapt their wording specifically to evade AI detection, making continuous evaluation necessary.
Another important metric is user behavior: whether receiving a warning actually reduces transfers, credential disclosures or other harmful actions.
Ultimately, the relevant outcome is prevented harm rather than classification accuracy alone.
··········
MOBILE HARDWARE, BATTERY CONSUMPTION AND APP INTEGRATION REMAIN DEPLOYMENT CONSTRAINTS.
The model has been optimized for typical smartphones in Singapore, according to Google DeepMind.
However, the announcement does not specify minimum RAM, supported processors, operating-system versions or storage requirements.
It also does not provide measured inference latency, energy consumption or thermal performance.
These details will determine whether the system can operate continuously without materially affecting the smartphone's usability.
The two-stage architecture is intended to reduce unnecessary computation, but the actual savings depend on how often messages trigger the more expensive analysis.
Long conversations may require additional memory and processing time.
Applications will also need mechanisms to determine which portions of a conversation should be supplied to Gemma.
Using too little context may conceal manipulation patterns. Using excessive context may increase latency and expose more private information to the local model than is necessary.
A further challenge concerns application integration.
A messaging provider can potentially integrate local scam detection into its own application, but a separate security app may face restrictions on accessing messages held by other services.
The announced architecture does not establish universal cross-application compatibility.
........
Deployment consideration | Current disclosure |
Supported operating systems | Not specified |
Supported smartphone models | Not specified |
Minimum RAM | Not specified |
Model download size | Not specified |
Inference latency | Not disclosed |
Battery consumption | Not disclosed |
Compatible messaging applications | Not announced |
Public installation method | Not announced |
Commercial pricing | Not announced |
........
Without these details, claims about broad smartphone compatibility should remain qualified.
The system has been developed and optimized for mobile deployment, but its actual accessibility will depend on future integration decisions.
··········
THE PROJECT EXTENDS SINGAPORE'S EXISTING MULTI-LAYERED ANTI-SCAM INFRASTRUCTURE.
Singapore already operates several mechanisms for detecting and disrupting fraud.
GovTech's Scam Analytics and Tactical Intervention System, known as SATIS, identifies suspicious websites and supports action against malicious online infrastructure.
ScamShield provides consumer-facing tools for identifying, reporting and blocking suspicious communications.
These systems address different parts of the fraud process.
Website analysis can identify malicious destinations. Telephone and message filtering can detect suspicious contacts. Financial institutions can monitor unusual transactions.
The new Gemma-based detector adds a potentially valuable layer: understanding manipulation as it develops within a conversation.
This is particularly relevant when an attacker does not immediately include a malicious link or request a payment.
However, a conversational model should not be regarded as a substitute for established fraud controls.
A scam may involve compromised accounts, authentic-looking websites, payment intermediaries and social engineering across several platforms.
No single classifier has visibility into every stage.
The strongest defensive architecture combines signals from messaging services, banks, telecommunications providers and relevant authorities while limiting unnecessary disclosure of personal information.
Singapore's approach illustrates how on-device AI could complement centralized threat intelligence rather than replace it.
··········
THE COLLABORATION COULD PROVIDE A REUSABLE MODEL FOR PRIVACY-PRESERVING FRAUD DETECTION.
The wider significance of the project extends beyond Singapore's immediate fraud problem.
Messaging services face a persistent conflict between protecting private communications and detecting increasingly sophisticated social-engineering attacks.
On-device language models offer one possible way to reduce that conflict.
A compact model can analyze relevant content locally while centralized systems continue to identify malicious domains, suspicious payment destinations and broader fraud networks.
The approach may also be adapted to local languages and institutions through targeted fine-tuning.
However, international deployment would require more than translating the Singapore model.
Fraud patterns differ between countries, as do consumer expectations, financial services, data-protection requirements and commonly impersonated organizations.
A model trained on Singaporean scam reports may not accurately recognize scams involving unfamiliar institutions or communication conventions.
Localization would require appropriate datasets, representative evaluations and mechanisms for responding to new fraud techniques.
The project's broader applicability will therefore depend on how efficiently the model can be adapted and validated in other environments.
It also remains unclear whether Google and GovTech intend to publish the fine-tuned weights, provide integration software or distribute the technology exclusively through selected applications.
Those decisions will influence adoption by messaging providers and public-sector organizations.
··········
THE NEXT MILESTONE IS A MEASURABLE CONSUMER DEPLOYMENT.
Google DeepMind and GovTech Singapore have demonstrated a technically credible architecture for analyzing potentially fraudulent conversations locally, using a lightweight screening model followed by a fine-tuned Gemma classifier.
The two-stage design addresses the computational limitations of smartphones, while local inference is intended to preserve the privacy of encrypted communications.
The project's use of anonymized local scam examples also reflects the importance of adapting fraud detection to regional institutions and social-engineering techniques.
Nevertheless, the October 14 announcement leaves several essential questions unanswered.
The partners have not identified the precise Gemma model, published detection metrics, disclosed hardware requirements or announced a general consumer release date.
Consequently, the system cannot yet be evaluated as an established, broadly available alternative to existing anti-scam applications.
The next meaningful evidence will come from controlled evaluations and real-world deployment, including measured false-positive rates, detection of previously unseen scams, battery consumption and reductions in financial harm.
The central innovation is the attempt to recognize manipulation across a private conversation without moving that conversation to a remote AI service. If the approach proves accurate, efficient and compatible with widely used messaging applications, it could become an important component of future mobile security systems.
For now, Google DeepMind and GovTech have established the architecture and research direction. Its effectiveness at protecting users remains to be demonstrated through published performance data and practical deployment.
··········
FOLLOW US FOR MORE.
DATA STUDIOS
datastudios.org
