Government intelligence agencies use a variety of tactics and strategies to combat terrorism. Nowadays, this mission has become a little more complicated given the unpredictability of terrorist behavior. Terrorists are increasingly cautious in their movements, especially in the use of social networks. To deal with this problem, several recent works have shown their interest in the identification of terrorist users in collaborative online environments. In this context, we propose a new framework for the detection of terrorist behavior by analyzing the relationships between individuals in a multidimensional network by applying a learning model (LR, CNN, SVM, BERT, KG-BERT, etc.), on multimodal data (texts, images, etc.) collected from several heterogeneous sources on terrorism. Our experimental results demonstrate the effectiveness of our framework with over 95% accuracy in detecting real terrorist behavior.