A Comparative Analysis of Deepfake Detection Using XceptionNet, MesoNet, and EfficientNet-B0 on UADFV Dataset

Deepfake media created with powerful deep learning methods are a big danger to the integrity of digital content. Reliable detection of such manipulated content has therefore become a critical research challenge. This paper presents a comparative analysis of three convolutional neural network architectures, namely XceptionNet, MesoNet, and EfficientNet-B0, for deepfake detection using the UADFV dataset. The models are evaluated using standard performance metrics including accuracy, precision, recall, F1-score, and computational efficiency. Experimental results demonstrate that EfficientNet-B0 achieves the highest accuracy (90.96%) and F1-score (91.23%), making it the most effective model for deepfake detection. XceptionNet is very close with an accuracy of 83.43%, while MesoNet, aimed for being computationally efficient, falls drastically with an accuracy of 52.30%. The finding provides practical insights into the trade-off between performance and efficiency while offering insights on the suitability of these models for real-time and resource-constrained applications.