Deep learning has been widely studied in many technical domains such as image analysis and speech recognition, with its benefits that effectively deal with complex and high-dimensional data. Our preliminary experiments show a high degree of non-linearity from the network connection data, which explains why it is hard to improve the performance of identifying network anomalies by using conventional learning methods (e.g., Adaboosting, SVM, and Random Forest). In this study, we design and examine deep learning models constructed based on Fully Connected Networks (FCNs), Variational AutoEncoder (VAE), and Sequence-to-Sequence (Seq2Seq) structures. For the extensive evaluation, we employ a broad range of the public datasets with unique characteristics. Our experimental results confirm the feasibility of deep learning-based network anomaly detection, with the improved performance compared to the conventional learning techniques. In particular, the detection model based on Seq2Seq with LSTM is highly promising, consistently yielding over 99% of accuracy to identify network anomalies from the entire datasets employed in the evaluation.

Deep learning has been given a great deal of attention with its success story in many areas such as image analysis and speech recognition. In particular, deep learning is good at dealing with high-dimensional data exhibiting non-linearity. Our preliminary study reveals a very high degree of non-linearity from network traffic data, which explains why it is hard to improve the detection accuracy by using conventional machine learning techniques (e.g., SVM, Random Forest, Ad-aboosting). In this study, we empirically evaluate deep learning to see its feasibility for network anomaly detection. We examine a set of deep learning models constructed based on the Fully Connected Network (FCN), Variational AutoEncoder (VAE), and Long Short-Term Memory with Sequence to Sequence (LSTM Seq2Seq) structures, with two public traffic data sets that have distinctive properties with respect to the distribution of normal and attack populations. Our experimental results confirm the potential of deep learning models for network anomaly detection, and the model based on the LSTM Seq2Seq structure shows a highly promising performance, yielding 99% of binary classification accuracy on the public data sets.

The Internet of things (IoT) is an emerging topic in a variety of industry fields such as manufacturing, engineering, automobile, etc. Especially, many automobile companies are competitively introducing a smart-connected car in conjunction with the IoT technology. For instance, some recent vehicles support autonomous driving using cameras and radar-based safety features such as pre-collision prevention system, lane departure alert with a steering assistant function, etc. Based on this example of the IoT-based smart connected car, we actually developed the camera-based blind spot assistance system. The power / serializer board and camera board consisting of a number of sensors, processors, and other hardware components were developed by our own system interface design, and we confirmed that the developed system successfully met quantitative evaluation criteria through testing. Furthermore, we are proposing improvement plans and an idea of object detection in the blind spot based on a deep learning methodology.

One of the recent news headlines is that a pedestrian was killed by an autonomous vehicle because safety features in this vehicle did not detect an object on a road correctly. Due to this accident, some global automobile companies announced plans to postpone development of an autonomous vehicle. Furthermore, there is no doubt about the importance of safety features for autonomous vehicles. For this reason, our research goal is the development of a very safe and lightweight camera-based blind spot detection system, which can be applied to future autonomous vehicles. The blind spot detection system was implemented in open source software. Approximately 2000 vehicle images and 9000 non-vehicle images were adopted for training the Fully Connected Network (FCN) model. Other data processing concepts such as the Histogram of Oriented Gradients (HOG), heat map, and thresholding were also employed. We achieved 99.43% training accuracy and 98.99% testing accuracy of the FCN model, respectively. Source codes with respect to all the methodologies were then deployed to an off-the-shelf embedded board for actual testing on a road. Actual testing was conducted with consideration of various factors, and we confirmed 93.75% average detection accuracy with three false positives.