Real-Time Ensemble-Based Deepfake Detection for Virtual Classrooms and Workspaces Using EfficientNet and Custom Attention-Enhanced CNN

The rising threat of deepfake technology poses serious challenges to the credibility of virtual classrooms and professional workspaces. As manipulated video content becomes increasingly realistic, real-time detection systems are essential to preserve trust in remote interactions. This paper presents a hybrid deepfake detection framework combining EfficientNet for global feature extraction and a custom Convolutional Neural Network (CNN) augmented with spatial and channel attention mechanisms for fine-grained facial analysis. An ensemble learning strategy fuses outputs from both models, and a predictive smoothing approach stabilizes real-time predictions across video frames. Experiments conducted on benchmark datasets such as FaceForensics++ and Celeb-DF demonstrate that the proposed system achieves superior performance with 97.5% accuracy and a 0.98 ROC-AUC score. Real-time evaluations show low-latency, frame-level detection at 30 FPS. These results validate the model’s reliability and adaptability. The proposed solution offers a strong foundation for real-time, robust deepfake detection in modern virtual communication environments.