STUDY OF MACHINE AND DEEP LEARNING ALGORITHMS FOR INTRUSION DETECTION IN IoT

Abstract

In this era of exponential internet boom IoT devices are also increasing with a rapid growth. This rapid growth also increases the risk of intrusion such as phishing at application layer, Dos & spoofing at network layer and node capture, malicious code injection & eavesdropping at physical layer. So, to prevent systems from these attacks it has become the desired need of time to implement an Intrusion Detection system model. In this paper we have briefly compared various global datasets and used most recent CSECICIDS-2018 dataset having 1.04 million of samples. We have implemented a Bi LSTM model having an Input layer, a reshape layer, two Bi-LSTM layers, a dense layer, a dropout layer and lastly an output layer. Proposed model is used for the prediction of a packet whether it is Benign and Not Benign using 11 important features 'Timestamp', 'Fwd Pkt Len Std', 'Fwd Pkt Len Mean', 'Fwd Pkt Len Max', 'Fwd Seg Size Avg', 'Pkt Len Std', 'Flow IAT Std', 'Bwd Pkt Len Std', 'Bwd Seg Size Avg', 'Pkt Size Avg', 'Subflow Fwd Byts' for training the model. This Bi-LSTM model have provided an accuracy of 99.554%, precision of 99.227% and F1 score of 99.612%. Further this model can be tested and improved on real-time intrusion scenario to provide improved results.