Comparative evaluation of lightweight CNN models for deepfake detection under social media compression conditions

The proliferation of deepfake media on social media platforms demands detection solutions that balance computational efficiency with robustness under realistic media conditions. While state of the art detection models such as XceptionNet and EfficientNet-B4 achieve high accuracy, their computational intensity limits real-time deployment at scale. This study presents a comparative evaluation of two lightweight convolutional neural network (CNN) architectures, MobileNetV2 (3.4M parameters) and ShuffleNetV2 (1.4M parameters), for deepfake detection under social media compression conditions. Both architectures were evaluated using 5-fold stratified cross-validation on a combined dataset of FaceForensics++ (C23 compression) and Celeb-DF v2, comprising approximately 13,529 videos and 295,976 face crops. Results demonstrate that ShuffleNetV2 outperforms MobileNetV2 across all key metrics (test AUC: 0.7839 vs 0.7674, accuracy: 86.17% vs 85.85%) while utilising 59% fewer parameters and training 23% faster. A transfer learning ablation study confirms that ImageNet pre-training is essential, as training from scratch yielded a near random performance (AUC: 0.5966). Compression robustness analysis at Q95 and Q50 JPEG quality levels reveals that both models maintain overall performance degradation below 3%, though ShuffleNetV2 exhibits a notable 5.33% AUC drop on FaceForensics++ at low quality. All models exceed real-time requirements by wide margins (>190 FPS at batch size 1 on A100 GPU). These findings establish lightweight CNN architectures as viable candidates for scalable deepfake detection and provide empirical evidence for selecting efficient architectures for social media deployment.