On-Device AI Scam Detection: Singapore's Privacy-First Defense

TL;DR: Google DeepMind and GovTech unveiled an on-device Gemma model at SICW 2026 to combat Singapore’s S$4 billion scam losses by processing data locally. This privacy-preserving approach bypasses encryption limits, ensuring user conversations remain secure while enabling real-time threat detection.

Key facts

• Google DeepMind and GovTech unveiled an on-device Gemma AI model for real-time scam detection at the Singapore International Cyber Week (SICW) on October 14, 2026.
• Singapore has suffered cumulative financial losses exceeding S$4 billion to scams since 2019, driven largely by sophisticated social engineering tactics.
• The system processes data locally using a two-stage pipeline: a lightweight classifier for screening and the fine-tuned Gemma model for in-context analysis of manipulative intent.
• Training occurred within GovTech’s Sovereign Sandbox on GCP, utilizing anonymized datasets that strictly excluded personally identifiable information to ensure privacy.
• In the first half of 2026 alone, Singaporeans lost approximately S$410.6 million to scams, representing a 17.9 percent year-on-year decline but remaining a critical threat.
• This initiative marks the first major technical deliverable of the National AI Partnership between Google DeepMind and the Singapore government under the Ministry of Digital Development and Information.

The S$4 billion gap

Scams have become Singapore’s most persistent crime problem, with cumulative financial losses exceeding S$4 billion since 2019 [1][4]. That number is not just a statistic; it represents the failure of traditional security tools to keep pace with increasingly sophisticated social engineering. To close this gap, Google DeepMind and the Government Technology Agency of Singapore (GovTech) have unveiled an on-device artificial intelligence model designed to detect scams in real time without compromising user privacy [2][5].

Announced at the Singapore International Cyber Week (SICW) on October 14, 2026, this initiative is the first major technical deliverable of the National AI Partnership between Google DeepMind and the Singapore government, established earlier that year under the Ministry of Digital Development and Information (MDDI) [1][2]. The core thesis here is clear: privacy-preserving security is no longer optional when criminals weaponize end-to-end encryption. By processing data directly on smartphones rather than transmitting messages to cloud servers, this system offers a new paradigm for digital defense.

How the model works

The system utilizes Google’s lightweight Gemma AI framework, fine-tuned specifically on local scam conversations and regional deception techniques [1][4]. Unlike traditional detection methods that rely on server-side scanning—a method rendered ineffective by end-to-end encryption on platforms like WhatsApp and Telegram—this new AI operates entirely on-device [5][6].

The architecture employs a two-stage pipeline for efficiency. Stage 1 uses a lightweight classifier to screen messages, while Stage 2 performs in-context analysis using the fine-tuned Gemma model to identify manipulative intent [4]. This approach allows the AI to flag suspicious interactions before financial or credential exposure occurs, all without sending user data outside the device [1][5].

Training took place within GovTech’s managed GCP environment known as the Sovereign Sandbox, using anonymized datasets that strictly excluded private or personally identifiable information [1][4]. Neither Google nor the Singapore government can access users’ conversations through this deployment model, ensuring that privacy is baked into the infrastructure rather than treated as an afterthought [1][4].

Why on-device matters now

The urgency of this solution is underscored by recent trends. In the first half of 2026 alone, Singaporeans lost approximately S$410.6 million to scams [2]. While this represents a 17.9 percent year-on-year decline, it remains a staggering figure driven primarily by investment fraud and impersonation of public officials [5]. Social engineering tactics now account for up to 80 percent of total financial losses from scams in the region [1][4].

As criminals increasingly leverage AI to generate convincing phishing content, Minister for Digital Development and Information Josephine Teo noted that the fight against scams has become harder than ever [3]. This new on-device model is a direct response to that challenge. It shifts the burden of protection from centralized servers—which can be bypassed or overwhelmed—to the individual device, where the conversation actually happens.

What comes next

The rollout will be phased in partnership with the Singapore government. While optimized for typical smartphones, it remains unclear if this technology will become a default feature of the Android operating system or when it will reach widespread consumer availability [5].

This project also coincides with broader national efforts to coordinate defenses against emerging threats, including quantum computing risks [3][6]. As digital scams grow more complex, the Singapore-Google DeepMind collaboration demonstrates that effective security does not require sacrificing privacy. Instead, it requires moving intelligence closer to the user.

Sources

1. On-Device AI Against Scams: Google DeepMind & MDDI (deepmind.google) — 2026-10-14
2. On-Device AI Against Scams: Google DeepMind & MDDI (deepmind.google) — 2026-10-14
3. Singapore and Google DeepMind Deploy On-Device AI That Detects Scams Without Cloud Access (www.techtimes.com) — 2026-10-14
4. Google, GovTech developing first-of-its-kind AI model to flag scams in real time for Android users (www.straitstimes.com) — 2026-10-14
5. Google DeepMind and GovTech Unveil On-Device Scam Detection Model (www.unite.ai) — 2026-10-13
6. GovTech and Google DeepMind developing real-time detection of scams on devices (www.channelnewsasia.com) — 2026-10-14