Reducing calculation requirements in FPGA implementation of deep learning algorithms for online anomaly intrusion detection

Abstract

Deep learning algorithms produced impressive results in the image and voice recognition fields. Machine learning approach can be implemented to improve anomaly detection method to detect novel attacks. We use dynamic fixed-point arithmetic to reduce Deep Belief Network (DBN) calculations in an FPGA. We trained a three-layer DBN using contrastive divergence with pipeline structure, fine-tuning the network using a softmax function. Our work using dynamic fixed-point arithmetic and pipeline structure reduced the calculation requirement of the DBN more than 30% compare to the 16-bit implementation. We used the MNIST dataset for evaluation before testing online intrusion detection and achieved accuracy of 94.6% on the NSL-KDD dataset and 95.1% on the HTTP CSIC 2010 dataset. We produced efficient resource utilization and detection speed of .008ms. Our design can be further improved to decrease deep learning resources during training and testing for online intrusion detection in low powered devices.