In computer security, Intrusion detection Systems (IDS) are mechanism of detecting anrnintruder in the system and notifying malicious activities to system administrator. Mostrnof IDS researches are on wired Local Area network (LAN) using KDD dataset. But thernwireless IDS needs its own research using dataset from wireless LAN. Since most of thernsecurity vulnerability features of wireless LAN is due of its nature and they are di er-rnent from wired LAN, wireless IDS needs to be studied independently from that of wiredrnLAN. The IDS researches on wireless LAN started recently. Until now there are somernresearch works like publishing Aegean Wi-Fi Intrusion Dataset (AWID) dataset publiclyrnfor the research community and evaluating the dataset using di erent machine learningrnalgorithms. But when we see the results from the previous research works, especially inrnthe case on Flooding and Impersonation attacks, it is clear that wireless IDS is not wellrnresearched and it needs further study for performance improvements.rnThe AWID dataset contains di erent data types which are numeric, string, and hexadec-rnimals. So before training the system and evaluation of its performance, the dataset isrnpreprocessed and nally 102 attributes are used for system training and evaluation. Alsorntwo stage feature selection is implemented to reduce the training cost and improve thernsystem performance by selecting the minimum number of most discriminant features.rnThe rst stage is removing duplicated attributes, which reduce the number of attributesrnin the dataset to 68. The second stage is done by applying Information Gain Ratiorn(IGR). Using three thresholds three dataset are prepared namely 41 attribute dataset, 34rnattribute dataset, and 25 attribute dataset to experiment the relation between number ofrnattributes in the dataset and the resulting system performance. The main classi cationrnsystem is implemented using Deep Belief Networks (DBN). Two stage training strategy is used to train DBN for classi cation. The rst stage is unsupervised pre-training using Re-rnstricted Boltzmann Machine (RBM) and the second stage is supervised ne tuning of thernpretrained DBN parameters using Back Propagation Neural Network (BPNN) algorithm.rnFinally after designing and implementing the system, a number of experiments have beenrndone to evaluate the system performance using di erent performance metrics. The sys-rntem was able to achieve 98.55% classi cation accuracy with 102 attributes and it wasrnable to improve this result to 98.97% with selected 34 attribute dataset evaluation. Butrnthe classi cation accuracy decrease to 98.74% while the numbers of attributes decreasernto 25. This shows that there is a limit in reducing the number of attributes and fromrnthe experiments it is found that the minimum number of the most discriminative at-rntribute list that was able to reach the maximum performance in the proposed systemrnis 34 attributes. The system has been tested also using 10-fold cross validation and itsrnclassi cation accuracy was improved to 99.96%.