Neural Network-Based Anomaly Detection in the Internet of Vehicles

The traditional anomaly detection algorithms such as the Kalman filter-based architectures have limited outlier detection in time series data streams with low magnitude. Such failure to detect outliers in data limits reliable data throughput. Advancements in IoT require the efficient transferring of large data sets without such anomalies. The report addresses this issue by further explaining the importance of efficient anomaly detection architectures, reviewing results of previous work in the field, and proposed architecture. A total of $\mathbf{1 5}$ previous studies on anomaly detection were examined, ranging from Kalman filter detectors to newer state-of-the-art deep neural network-based architectures. Some of the reviewed architectures can be grouped because of similar theories and a slight variation. Each study discussed has its advantages and limitations, all listed in their respective summary. After reviewing previous work and building architecture, the proposed method was obtained to remove previously mentioned limitations. Many different methods, including simple DNN based AE and LSTM/GRU-based methods, have been implemented. This paper implements bidirectional LSTM AE and introduces the usage of bidirectional GRU-based AE for anomaly detection. The F1-Scores of these two methods are about 4% higher than the traditional methods and outperform the architectures implemented previously on the same ECG dataset.