Addressing Video-Replay-Attacks in Active Virtual Sports: A New Dataset and Deep Learning-Based Approach

Video-replay attack detection is an important task in active virtual sports that aims to differentiate between legitimate captures and replay attacks. This study presents a novel dataset of videos depicting users performing football juggling tasks, along with corresponding videos simulating replay attacks using multiple monitors and smartphone cameras. To evaluate the proposed dataset, we developed a deep learning approach based on the EfficientNet architecture and performed 5-fold cross-validation experiments. Results show that the proposed approach achieved an average accuracy of 99.1% on the presented dataset. Additionally, an ablation study was conducted to assess the significance of various dataset aspects, including diversity.