Abstract: In order to solve the problem that data samples for recognition of traffic terrorist attacks are too limited to meet the demand of DNN for massive training data, based on the data-enhanced deep learning, a terrorist attack recognition approach was proposed. Firstly, data enhancement strategy and method were designed to gain enough training samples. Secondly, cost function was calculated via both original and enhanced training samples during network training process to overcome overfitting and improve robustness. Then, DNN was developed by using stacked multiple sparse auto-encoder(SAE) and classification layers. Moreover, feature extraction and pattern recognition were integrated via layer-by-layer self-learning and supervised and reversed fine-tuning of enhanced samples in order to accurately identify attack behaviors. Finally, the proposed approach was illustrated and demonstrated with attack data from global terrorism database(GTD). The results show that the feature extraction ability and recognition results of the enhanced deep learning algorithm have been improved compared with comparison algorithm with limited data samples, and its average recognition accuracy can reach as high as 98.75%.

Key words: limited samples, traffic terrorist attack, behavior recognition, data enhancement, deep neural network(DNN)