Singapore and DeepMind Announce Mobile Anti-Scam AI

Facing losses exceeding 4 billion Singapore dollars since 2019, GovTech and Google DeepMind unveil a device-executed scam detection model. Fine-tuned on local conversations and tested under energy constraints, it promises real-time alerts without extracting encrypted messages. The collaboration was presented at SICW 2026 with guarantees of sovereignty and data anonymization.

The solution respects encryption while protecting users

The collaboration was showcased during the Singapore International Cyber Week 2026. The partners present it as an example that reconciles user protection and end-to-end encryption for messaging platforms. The model was refined in a GCP environment managed by GovTech, with isolation of training data. No private data or personally identifiable information from citizens was used; the training sets were anonymized and derived from scam reports and intelligence feeds. The model underwent stress testing on synthetic dialogues and local corpora covering various categories of scams. The project illustrates cooperation between governments and model providers to develop technologies of public interest.

Real-time on-device detection with a two-step pipeline

The model processes conversations directly on the device to preserve privacy. This choice of local analysis allows for quicker detection of manipulative intentions, enabling action before any financial or identity data exposure. The goal is real-time alerts on targeted and local attempts without compromising privacy. To balance performance and autonomy, the architecture adopts two steps: a lightweight classifier first filters incoming messages with low energy consumption, followed by contextual evaluation via the refined Gemma analyzing a broader portion of the exchange to confirm the threat or lift a false alert. Energy efficiency is explicitly identified as a key constraint on smartphones.

Warnings guiding the user before any sensitive sharing

When a manipulation scheme is detected, a local notification appears on the user's phone. It prompts the user to pause and exercise caution before transmitting sensitive information or funds.

A model rooted in the Singaporean context

The lightweight Gemma model has been fine-tuned on scam conversations specific to the local context. The training covers impersonations of public institutions, banks, and local figures, as well as recurring regional scams. This approach aims to improve the distinction between ordinary social exchanges and manipulation attempts. The optimization targets typical smartphones in the country and takes into account that many messaging services are end-to-end encrypted, rendering server-side analysis ineffective.

A technical response to massive financial harm

In Singapore, scams account for more than half of reported crimes, with cumulative losses exceeding 4 billion Singapore dollars since 2019. Modus operandi have evolved from simple phishing to more elaborate social engineering scenarios, with these attacks accounting for up to 80% of losses. In this context, GovTech and Google DeepMind collaborate under a National AI Partnership that has resulted in a device-executed detection model.

Towards integration by messaging platforms and operating systems

Messaging platforms and operating system providers can directly integrate lightweight models of this type to offer local protection. This integration aims to secure conversations without exposing private data beyond the device.