Disrupting the Scam Cycle: AI for Safer Communications

Abstract

Fraudulent phone calls are among the fastest-growing cybersecurity threats worldwide. In India alone, losses crossed 11,333 crore rupees in 2024. To address this, the paper introduces Live Scam Call Shield, a hybrid system that combines on-device machine learning with optional cloud support to deliver real-time scam detection while preserving user privacy. The system integrates directly into Android’s default dialer via native CallReceiver mechanisms, ensuring seamless operation without requiring users to switch apps or rely on external APIs during active calls. In testing, our model achieved 95.7% accuracy and a 94.8% F1-score on a dataset of 12,847 calls (85 hours of audio), with a median detection latency of just 4.2 seconds. A secure REST API backend supports encrypted model updates and crowdsourced feedback, while keeping privacy at the core. Field trials with 75 beta users across fraud-prone regions (Maharashtra, Delhi, and Karnataka) confirmed 96.2% precision with only 1.4 false positives, significantly reducing unnecessary alerts compared to existing solutions. Our findings show that privacy-first, on-device detection can rival cloud-based systems (95.7% vs. 99% accuracy) while offering offline resilience and greater user control—critical for regions with limited or unstable internet access. This work delivers the first production-ready solution that blends local intelligence with optional remote enhancement, setting a new benchmark for secure, privacy-preserving fraud detection in telephony.