Advancing the Intrusion Prediction Model through a Hybrid Framework

This study investigates the application of Artificial Intelligence models to enhance intrusion detection and prediction in high-security environments by leveraging CCTV footage for advanced surveillance. Using a dataset that captures diverse intrusion behaviors, we implemented a robust feature extraction pipeline that combines MediaPipe and YOLO, and applied various deep learning architectures, including CNN, RNN, and LSTM, to evaluate their performance. The primary objective was to develop predictive models that outperform traditional detection systems, thereby enabling proactive security measures. Experimental results showed that the LSTM model, configured with a window size of 20 frames, achieved the highest accuracy of 98.92%, outperforming other models across multiple evaluation metrics. To assess real-world applicability, we compared the inference times of the models in both GPU and CPU environments. The 1D-CNN model with a 15-frame input demonstrated the fastest prediction speed in both settings. While this research marks significant progress in intrusion prediction, future work should focus on expanding dataset diversity, incorporating multimodal inputs such as audio and video, and optimizing models for real-time applications. These efforts are expected to further improve AI-based security systems, transitioning them from reactive to proactive intrusion management.