Single-layer KAN for deepfake classification: Balancing efficiency and performance in resource constrained environments

Abstract

Deepfakes, synthetic media created using artificial intelligence, threaten the authenticity of digital content. Traditional detection methods, such as Convolutional Neural Networks (CNNs), require substantial computational resources, rendering them impractical for resource-constrained devices like smartphones and IoT systems. This study evaluates a single-layer Kolmogorov-Arnold Network (KAN) with 200 nodes for efficient deepfake classification. Experimental results show that KAN achieves 95.01% accuracy on the FaceForensics++ dataset and 88.32% on the Celeb-DF dataset, while requiring only 52.4 MB of memory, 13.11 million parameters, and 26.21 million FLOPs, significantly less than state-of-the-art CNNs. These verified metrics highlight KAN’s potential for real-time deepfake detection on edge devices. Untested capabilities, such as robustness against adversarial attacks, are proposed for future research. This work aligns with the United Nations Sustainable Development Goals, specifically SDG 9: Industry, Innovation, and Infrastructure and SDG 16: Peace, Justice, and Strong Institutions.