Google DeepMind and Singapore build an on-device Gemma scam detector for Android

What did Google DeepMind and Singapore announce?

Google DeepMind and GovTech said on 14 October that they have built an on-device scam detection model designed for typical smartphones in Singapore. It was presented at Singapore International Cyber Week, the same event where the government launched the national cybersecurity strategy we covered on 13 October. In a Google DeepMind blog post by JJ Geewax and Miguel de Andrés, the companies describe it as a way to catch scams inside private conversations without sending those conversations to a server.

The model looks for manipulative patterns in chats. When it finds them, it shows a warning on the phone telling the user to pause before sharing sensitive information or sending money. According to The Straits Times, it is meant to work across messaging apps such as WhatsApp and Telegram.

The work sits under Google DeepMind’s national AI partnership with Singapore, announced in May 2026, which already covers health research, education and a Gemma-powered running assistant for blind athletes. Minister for Digital Development and Information Josephine Teo cited the scam project in her SICW speech. “If our devices can recognise signs of manipulation early, we can help potential victims avoid financial or credential loss, while keeping their personal messages on their devices,” she said, as reported by CNA.

How does the on-device scam model work?

Google says generic models often miss scam tactics that are specific to a place. So the team took Google DeepMind’s lightweight Gemma model and fine-tuned it on local scam conversations. The examples ranged from impersonations of public institutions, banks and local figures to scams trending in the region. The aim, the blog says, is to tell ordinary social chatter apart from real manipulation.

Running a classifier on every message would drain a phone’s battery, so the system works in two steps. Google describes the pipeline this way:

• Stage 1: a lightweight, efficient classifier screens incoming messages using minimal power, looking only for potential threats.
• Stage 2: the fine-tuned Gemma model reads a larger snapshot of the conversation to confirm the threat or discard a false alarm.
• Output: if a threat is confirmed, a real-time warning appears on the device advising the user to pause before sharing information or funds.

How is privacy handled?

Most scam filtering today happens on servers, where a provider scans messages in the cloud. That does not work for apps with end-to-end encryption, because the provider cannot read the messages. Google’s answer is to move the analysis onto the phone, where the conversation is already decrypted for the user.

Google DeepMind says the model will only be able to access on-device conversations with the user’s consent. The Straits Times quotes the blog’s authors as saying that because the model runs entirely on the device, “neither the Singapore Government nor Google will be able to access users’ conversations through its deployment.”

The companies also list three governance rules for how the model was built:

• Sovereign sandbox: fine-tuning took place in a Google Cloud environment managed by GovTech, keeping the training data isolated.
• No citizen data: training excluded private or personally identifiable information and used anonymised datasets derived from scam reports and intelligence feeds.
• Stress-testing: the model was tested against synthetic dialogues and local text collections covering different scam types and impersonation tactics.

Why is Singapore going after conversation scams?

Scams make up more than half of Singapore’s reported crimes and have cost victims more than S$4 billion since 2019, according to Google DeepMind. The Straits Times reports S$410.6 million was lost in the first six months of 2026 alone, with e-commerce, phishing, investment, job and government-official impersonation scams the most common.

The type of scam matters here. Google says tactics have shifted from simple phishing links to social engineering, where a scammer builds a convincing conversation before asking for money, and that these scams now account for up to 80% of losses. A link filter cannot see that kind of fraud; something that reads the conversation might.

Teo said AI has made the fight “harder to win”, but argued against retreating from the technology. “Our response should not be to turn away from AI because scammers misuse it. Instead, we must tune up our defences and make them better,” she said. She pointed to an existing GovTech and police system that uses AI to scan more than 600,000 websites a day and disrupts about 20,000 malicious sites a month. On-device detection adds a check at the point where scam conversations actually happen.

What else did Google offer Singapore’s cyber defenders?

Google made a second announcement the same day. It said it will give the Singapore government access to its most advanced AI cybersecurity tools to help defenders find, verify and fix vulnerabilities in minutes. “By proactively patching weaknesses ahead of time, the program aims to protect essential systems and services from operational disruption and cybersecurity threats,” Google said, according to The Straits Times.

Google did not name the program in the material we read. Google already runs a limited-access scheme for governments and critical infrastructure called the Fairwind Program, which gives approved partners a cyber-defence version of Gemini 4 Argon and its CodeMender patching agent. Whether Singapore’s access runs through Fairwind has not been confirmed, so treat that link as context rather than fact.

The two announcements fit Singapore’s strategy. As we reported on 13 October, the government has said it will not build its own frontier models and will instead build the software around the best available ones. A scam detector built on Google’s open Gemma family, fine-tuned inside a GovTech-run environment, is an example of that approach.

When can people use it?

Not yet. Google has not announced a release date, and The Straits Times reports it is unclear whether the detector will become a default Android feature. CNA reports Google said more details on how users can install and use the model will come soon.

Google DeepMind says messaging platforms and operating system makers can integrate the lightweight models directly, and that it is exploring similar deployments in other countries. The model will also be retrained continuously as scam tactics change. On-device Gemma work is moving quickly at Google; last week it released EmbeddingGemma 2 for local search, and Android has described a hardware-isolated vault for on-device AI. How the scam model will be packaged and protected on phones has not been described.

What are the limits and open questions?

The announcement is light on numbers. Google has not published detection rates, false-positive rates, the size of the Gemma model, which phones it supports, or how much battery it uses. It says the model was stress-tested, but has not shared results. Until it does, “real-time scam detection” is a claim about design rather than measured performance.

Consent and scope also need detail. A tool that reads private chats, even locally, raises fair questions: how consent is asked for and withdrawn, whether the warnings or any metadata leave the phone, and who reviews mistakes. Google’s statement that neither it nor the government can see conversations through the deployment is a strong commitment, and readers should watch for a technical description or independent audit that backs it up.

Attackers will also adapt. Any AI that reads untrusted messages can be targeted with text written to fool it, a risk our explainer on prompt injection covers. Google says the model will keep being retrained as tactics change. Singapore also has four official languages, English, Malay, Mandarin and Tamil, and Google has not said which of them the model handles. A warning only helps if it is accurate and people heed it, and that will only be clear once it ships to ordinary phones.