Deep Learning Based Classification of IoT Data Sets

Abstract

Classification refers to identify different classes of labeled or unlabeled data on the basis of features, distance between different features and similarity between different data objects. The classification of IoT data means the identification of classes from network data which is generated through different IoT devices and sensors. Therefor it can be said that classification of network traffic is also a network intrusion detection system in which intruders classes and normal data are classified. Furthermore different classes of attacks can also be identified in network intrusion detection system. Thus an Intrusion Detection System (IDS) refers to monitoring network data facts, information, rapidly identifying intrusive behavior and avoiding damage caused by an intruder. This also classifies them according to their features. Traditional intrusion/intruders detection techniques primarily focused on rule based methods and data mining approaches but they are at a disadvantage because they are unable to detect new types of attacks and are also slow to detect them. Although Deep Learning (DL) has successfully applied to many Machine Learning (ML) and Data Mining (DM) related issues, however less work has been done in the domain of IDS, thus there is a room available for research in this domain. As a result IDS has become an important layer in all the state-of-the-art ICT systems due to a craving towards cyber security and protection in the day-to-day world. So variety of challenges are being faced nowadays of network intrusion which are continuously increasing. These are due to vulnerabilities in software, hardware and network protocols. Therefore, stronger IDS is required; ML and DM have further strengthened the IDS technology. At the same time threat has also become more sophisticated. Now the stronger IDS are required to classify network traffic with best performance. Presently, data classification has become a great challenging task. Many esteemed ML techniques (supervised and unsupervised) are used to overcome this situation. In this thesis, DL techniques implemented for packet inspection and attack identification. Now over fitting and structured optimization techniques are used in IDS. In this thesis, we have proposed a Deep Neural Network (DNN) based IDS. DL based system monitors the traffic coming from authentic and non-authentic sources. It classifies and segregates malicious traffic with accuracy up to 99.78% for mixed data type and 99.91% for numerical data type. KDD99 data set was used for experimentation and comparative analysis with previous techniques which shows encouraging results. Feature Extraction techniques are also used in this thesis but did not achieved desired accuracy with feature reduction. Proposed model achieved highest accuracy with all features.