메신저 사기, 클라우드 말고 폰 안에서 잡는다

Sources — ondevice-scam-gemma-20261015-1252

Official

·https://deepmind.google/blog/on-device-ai-scam-detection-mddi/

Google DeepMind blog, Oct 14, 2026: "Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection"

Authors: JJ Geewax and Miguel de Andrés

Presented at Singapore International Cyber Week (SICW 2026)

Verified claims used in slides/caption

·Fine-tuned lightweight Gemma on local scam conversations (impersonation of institutions/banks/figures, regional schemes)

·On-device: neither Singapore Government nor Google can access users’ conversations through deployment

·User consent required when integrated into apps

·Scams = over half of Singapore’s reported crimes; more than S$4 billion losses since 2019; social-engineering scams up to 80% of losses

·Two-stage pipeline: (1) lightweight classifier screens messages, (2) fine-tuned Gemma analyzes larger conversation snapshot

·Real-time on-device warning to pause before sharing sensitive info/funds

·Phased rollout with Singapore Government; exploring similar tech for other countries

Uncertain / limits (stated clearly)

·Not a Korea consumer toggle yet — phased Singapore apps first; other countries exploratory

·Exact app list / public consumer availability date in Singapore not specified in the blog