Real-Time Deepfake Detection on IoT Edge Devices via Lightweight CNN and WebRTC Integration

Abstract

Generative AI has made deepfake injection into live video conferencing an accessible, low-cost attack vector, yet no prior work has demonstrated end-to-end deepfake detection running entirely on commodity IoT edge hardware within an active WebRTC session. This paper closes that gap with four novel contributions. First, we design a hardware-adaptive dualbackend inference engine that selects at runtime between a 3 MB post-training quantized TFLite MobileNet CNN (Raspberry Pi 5, CPU-only) and the apple/mobilevit-xx-small Vision Transformer (NVIDIA Orin NX, GPU), enabling a single codebase to span two orders of magnitude in compute capability without cloud offloading. Second, we introduce a rolling-window enforcement mechanism, a 75-frame deque requiring a sustained 70% fake ratio before peer disconnection, that decouples perframe classifier noise from enforcement decisions, achieving a false positive rate below ${10} {\%}$ and end-to-end detection latency under 30 seconds in live testing. Third, we provide empirical resource profiling of both platforms under live Web Real-Time Communication (WebRTC) load, confirming that TFLite inference sustains $\geq 5$ frames per second (FPS) at ${12 - 16 {\%} C P U}$ utilization and a 55.7 °C peak temperature on the Raspberry Pi 5, well below the ${80}^{\circ} \mathrm{C}$ throttle threshold, and that a dedicated CUDA worker thread eliminates unbounded GPU memory growth on the Orin NX. Fourth, we contribute a root-cause analysis of domain shift and catastrophic forgetting encountered when deploying MobileViT on live webcam feeds, establishing that quantized CNNs are the superior production choice for edge deepfake detection under current dataset constraints and providing practitioners with concrete remediation guidance for ViT fine-tuning failures.