Real-Time Multi-Modal Deepfake Detection for Edge Devices Using Lightweight CNNs

Abstract

The rapid-fire increase of generative artificial intelligence has led to the wide creation of largely realistic manipulated images, vids, and audio, generally known as deepfakes. While being deepfake discovery styles have shown promising results, numerous depend on computationally high-end models and pall- grounded processing, limiting their felicity for real- time and sequestration-sensitive operations. This paper proposes a feather light multi-modal deepfake discovery frame designed for real- time working on edge bias. The system integrates visual and audio analysis using effective CNN- grounded models and applies point- position emulsion to find cross-modal inconsistencies. Preprocessing ways similar as face- region birth, videotape frame slice, and Mel- spectrogram generation are recommended to regularize different media types. The proposed frame is optimized for CPU- grounded prosecution and performs all conclusion locally, barring pall reliance. Experimental evaluation on standard datasets demonstrates that this approach provides dependable deepfake discovery while maintaining low quiescence, making it suitable for practical edge- grounded media authenticity verification for privacy and security.