Abstract
<jats:p>Deepfake images generated using advanced generative artificial intelligence techniques have become increasingly difficult to distinguish from authentic images, creating significant challenges for digital security, identity protection, and misinformation prevention. Although deep learning-based detection models achieve high accuracy, many require substantial computational resources, limiting their deployment on mobile and edge devices. This paper presents a comparative analysis of eleven activation functions within the lightweight MobileNetV2 architecture for efficient deepfake image detection. Rather than modifying the network architecture, the activation layers in the inverted residual bottleneck blocks are replaced with ReLU, ReLU6, Leaky ReLU, PReLU, ELU, SELU, CELU, Swish, Hard-Swish, Mish, and GELU while maintaining identical training settings. Experiments are conducted on the 140K Real and Fake Faces dataset and the GRAVEX-200K dataset under both in-domain and cross-domain evaluation protocols. The results demonstrate that activation function selection significantly influences feature learning and generalization performance. In in-domain evaluation, CELU achieved the highest accuracy of 99.84% on the 140K dataset, while Mish achieved 96.82% on the 200K dataset. Cross-domain experiments further reveal that models trained on the larger and more diverse GRAVEX-200K dataset exhibit substantially stronger generalization, with hyperparameter settings achieving up to 99% accuracy and 0.9939 F1-score, where Swish, Hard-Swish, ELU, and ReLU6 consistently delivered superior performance. The findings demonstrate that optimizing activation functions provides a simple and computationally inexpensive approach to improving lightweight deepfake detection without increasing model complexity, making MobileNetV2 more suitable for deployment on resource-constrained edge and mobile devices.</jats:p>