Web Systems which are the backbone of information resources, communications, and personal information management, attackers might take advantage of their vulnerability and beguile them to get access to sensitive data or the web servers and apps in full. Wider usage of the Internet and its features have come a long ago, with both advantages and disadvantages. The security and its side effects have been discussed by researchers in various aspects of the network. HTTP, one of the most widely used network protocols, has paid a huge price due to various intrusions like SQL injections, code injections, and cross-site scripting (XSS). To handle these intrusions, various intrusion detection algorithms have been proposed and addressed in the literature. The accuracy and timeliness of these detections has been an issue due to false positives and the amount of information that ought to be processed and delivered within a short span of time. In recent times the classic classifiers have become outdated and detecting abnormal traffic in a web system has become a hassle. In this research, we propose handling intrusion detection over a Web System using hybrid Deep Learning (HDL) based classification and robust feature extraction using a modified chicken swarm optimization algorithm (MCSO). This approach also incorporates the idea of deep learning, which makes it possible to have a peer-to-peer learning system for aberrant patterns with a fewer number of characteristics, hence reducing the amount of time needed to complete the work. In order to distinguish or classify data that the web system has to deal with, a hybrid computational intelligence-based classifier algorithm is used. The combination of the hybrid classifier is fuzzy neural network with Long Short Term Memory (LSTM) which is basically used to classify the attacks and distinguish between normal and anomalous data. The use of helps in understanding the nature of intrusions over time, which is constant and predictable, On the other hand, with the assistance of deep learning and a method called feature extraction, which pulls important information from noisy data, we can do this. Lastly, the findings of the experiments show that this technique has an excellent detection performance, with an accuracy rate that is more than 98.7%.