Semantic-Aware Lightweight AI Model for Deepfake Image Detection in Online Retail Platforms

Deepfake detection in e-commerce platforms demands lightweight, efficient, and accurate models capable of real-time performance. This paper proposes a semantic-aware, lightweight ShuffleNet-based model optimized for detecting image-based deepfake content. The proposed model integrates ShuffleNet with a Semantic Knowledge Graph (SKG) for enhanced deepfake image detection. The SKG links extracted visual features with contextual metadata, enabling a more interpretable and knowledge-driven classification process. The proposed model achieves an accuracy of 76.15%, precision of 80.97%, recall of 76.15%, and F1-score of 74.79% while significantly reducing computational costs. Compared to standard architectures like DenseNet, MobileNet, and EfficientNet, the model achieves the lowest FLOPs (295.57M) and parameter count (1.26M). These results highlight the model's ability to outperform existing architectures in balancing performance and computational efficiency. The proposed solution is ideal for real-time, resource-constrained environments, positioning it as an effective tool for combating deepfake challenges in online retail.