Enhancing Software Security with CNN-LSTM Hybrid Model: Vulnerability Detection

In today’s era, due to the rigorous growth in information and communication technologies the software plays a vital role. The programming languages are considered to be essential for our professional/work activities. Amongst them the widely used are the high-level programming languages. Due to new way of programming method, it's easy to code and understand the syntax and make the beginner user-friendly. The programming language covers various areas like web programming, artificial intelligence, deep learning and data science. The rapid growth in usage programming leads in to huge explorer, however this makes the languages more vulnerable. Vulnerability of programming language is flaw, patch or weakness through which the critical information or data can be attacked. Python is the beginner-friendly programming language which is dynamic and has third party libraries. It creates unique circumstances for identifying problems. It addresses various issues in the development of software and implementations are dependent on one another. To detect the software vulnerabilities in the Python programming language the proposed system uses layered fusion deep learning method, multiple convolution layers for local feature extraction and Long Short-Term Memory (LSTM) model to frame the dependency in the source code. The packages like TensorFlow, Keras and PyTorch, are python- based functions used during the implementation process. It’s necessary to apply an attention mechanism which allows valuables properties of the code to increase the performance of identification. To decrease the training time and improve the generalization the transfer learning method is applied. The difference is very clear as compared with conventional test and the proposed method. The model is adaptable and it can integrate to large code mass and project -programming languages, it will be the right thing to apply in real world as far as security of the software is concerned. The proposed model achieves an accuracy 94 %, precision 93%, recall 96% and F1 score of 94%.