Automation in the complex environment of modern electronic warfare is a critical and necessary issue in electronic identification and support systems to detect real-time and accurate threat radars. These systems search, discover, analyze, and identify the parameters of radar signals. One of the key issues in these modern innovative systems is pulse repetition interval (PRI) modulation detection. However, detecting PRI modulation is challenging in the real environment due to destructive factors such as missed pulses, unwanted pulses, and large outliers (caused by antenna scanning) that lead to a noisy sequence of PRI variation patterns. This article examines the effects of destructive factors on detecting PRI modulation in radar signals using two convolutional neural network models, VGG 16 and LeNet 5, with two different structures. The paper uses simulations based on the actual environment to generate data and consider malicious agents with various percentages. The number of images obtained by applying the sum of malicious agents on it for each range of malicious agents (with different percentages) considered is 30,000 images for 6 Modulation type is standard. Then, the VGG16 model is trained using the transfer learning method, and the LeNet 5 model is trained using the zero training method. The simulation results show that the accuracy of training and testing the models decreases significantly with the increase in the percentage of destructive factors. Also, destructive agents' effects on models' performance have been investigated. The results have shown that LeNet 5 is more resistant to malicious agents and maintains more accuracy. Finally, this analysis indicates that to choose the right model for electronic identification and support systems, the changes caused by malicious agents should be provided according to these factors, and appropriate strategies should be applied.