Intrusion detection system for wireless sensor networks based on Deep learning

Abstract

Human kind is an evolving creature that has always looked to make his life easier, that led him to create modern civilization. Technology is a non-stop arrow that keeps moving forward towards a bright future, brilliant innovations and new challenges. Wireless sensor networks (WSN) is one of the leading topics in our modern society, it’s the key to achieve the dream of global smart cities with IOT (Internet Of Things ) devices and it’s the key for many exciting domains like remote medicine, smart agronomy, etc. However, WSNs are exposed to cyber threats every instant, whether they are intentionally by an adversary or because of poor system management. Thus, security in WSN is very critical and challenging at the same time due to the constraints of sensors like the concise abilities such as limited memory, energy consumption constraint. Therefore, any security measurements must take into consideration these factors in order to grant the network the fullest performance that it should get without any delay or packet loss or any sort of abnormal functioning. In this thesis, an Intrusion detection system (IDS) engine is proposed based on machine learning and deep learning techniques combined with features engineering to obtain the highest results for good packet classification. KDDCUP99 and NSL-KDD are the datasets used to test the model, the data is preprocessed and then passed as input to a feature selection algorithm XGBoost that selects columns based on features scores .Since WSNs computation power is limited, binary classification is used, to classify normal from abnormal traffic using a Deep Neural network of four layers, input layer, two hidden layers and the output layer. The results obtained during this study are accurate and precise compared to what researchers have accomplished in their published papers. For KDDCUP precision and accuracy scores are 99.86%, for NSL-KDD 83.21% precision score and 76.21% accuracy score.