Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection

On‑Device AI for Scam Detection: A Singapore‑Google DeepMind Collaboration

Scams—ranging from investment fraud to fake job offers—account for more than half of reported crimes in Singapore and have caused over S$4 billion in losses since 2019. As attackers shift from simple phishing to sophisticated social‑engineering conversations, traditional server‑side detection struggles to keep pace while preserving user privacy. In response, Google DeepMind and the Government Technology Agency of Singapore (GovTech) have partnered under the National AI Partnership to fine‑tune a lightweight, on‑device model that flags manipulative dialogue without exposing personal messages to the cloud.

---

Why On‑Device Detection Matters

• Privacy First – The model runs entirely on the user’s device and only accesses conversations with explicit consent. Neither Google nor the Singapore government can retrieve raw chat data, aligning with end‑to‑end encryption standards.
• Speed and Context – Local processing eliminates latency associated with server round‑trips, enabling real‑time warnings that appear instantly as a suspicious conversation unfolds.
• Battery Efficiency – A two‑stage pipeline balances power consumption: a lightweight classifier screens inbound messages, and a more detailed analysis using the fine‑tuned Gemma model runs only when the first stage flags potential risk.

---

Building a Model Tailored to Local Threats

Generic AI models often miss region‑specific deception tactics. To address this, DeepMind fine‑tuned its compact Gemma model on anonymised, locally sourced scam dialogues—impersonations of public institutions, banks, and trending regional schemes. Training took place within GovTech’s sovereign sandbox on Google Cloud Platform, ensuring that no personally identifiable information entered the process. Synthetic dialogues and curated corpora were used for stress‑testing, covering a wide spectrum of evolving impersonation techniques.

---

How the Detection Flow Works

1. Stage 1 – Efficient Screening

- A minimal‑footprint classifier scans incoming messages for keywords and patterns indicative of scams. - This stage consumes negligible power and runs continuously in the background.

1. Stage 2 – Contextual Confirmation

- When Stage 1 raises a flag, the fine‑tuned Gemma model examines a broader snippet of the conversation. - If manipulative intent is confirmed, the system triggers an on‑device alert, prompting the user to pause before sharing sensitive information or funds.

The alert appears directly within the messaging app, preserving the user’s workflow while delivering a clear, actionable warning.

---

Governance and Transparency

The collaboration was showcased at Singapore International Cyber Week (SICW 2026), highlighting a governance framework that includes:

• Sovereign Sandbox – All training occurred in GovTech‑managed infrastructure, keeping data isolated from external networks.
• No Citizen Data – Only anonymised scam reports and intelligence feeds were used; private messages never entered the training set.
• Rigorous Stress‑Testing – Synthetic and local corpora ensured robustness against emerging scam variants.

These safeguards demonstrate that high‑impact AI can be developed responsibly, respecting both security and privacy.

---

Looking Ahead

The on‑device scam detection model sets a precedent for how governments and AI providers can co‑create public‑good technology. While the rollout will be phased, the approach offers a template for other nations seeking to protect citizens without compromising encrypted communications. Future work may expand the model’s capabilities to other threat vectors—such as deep‑fake audio or fraudulent links—while maintaining the same privacy‑first architecture.

---

Bottom line: By embedding a fine‑tuned, privacy‑preserving AI model directly onto smartphones, Google DeepMind and GovTech Singapore are delivering real‑time scam protection that respects user confidentiality, paving the way for safer digital interactions worldwide.