Comparative evaluation of lightweight CNN architectures for efficient and explainable deepfake detection

Abstract

Deepfake images generated using the advanced generative models have raised serious concerns regarding digital media authenticity, misinformation, and security. Detecting such manipulated content is essential, in a real-world scenario where computational resources may be limited. This work presents a comprehensive study of lightweight Convolutional Neural Networks (CNN) architecture for images based on deep-fake detection. Several efficient models including EfficientNet-B0, MobileNet-V3 (Small), ShuffleNetV2, SqueezeNet and ResNet-18 are evaluated to analyze their performance, computational efficiency and robustness. These models are trained and tested on large scale OpenForensics: Wild Dataset with standard preprocessing and regularization techniques. In addition to conventional performance, metrics such as accuracy, precision, recall, and F1-Score computational complexity are assessed through the parameter counts and Floating Point Operations (FLOPs). Furthermore, Explainable Artificial Intelligence (XAI) Scores calculated using Deletion Area Under the Curve (AUC) and Insertion AUC are also reported. The experimental results demonstrate competitive deepfake detection performance while maintaining significantly lower computational cost. The finding highlights the potential of efficient deep learning models for practical and scalable deep-fake detection systems, while improving interpretability and robustness.