Anomaly Detection in IoT Data Streams Based on Long Short-Term Memory

Nowadays, the swift growth of Ambient Intelligence (AmI) in Internet of Things (IoT), sensing devices have been generating many data streams in intelligent scenarios. The IoT deployment issues, complexity of systems is leading to abnormal behavior in data streams. Moreover, the dynamic and consecutive data streams generated in IoT systems leads to difficulty in determining anomaly detection. Therefore, this paper presents a Long Short-Term Memory (LSTM) and Sparse Autoencoder method for classification of data to determine anomaly behavior in IoT data streams. Firstly, a min-max normalization technique is employed to scale the input data for better results and Sparse Autoencoder utilizes sparsity penalty to capture relevant features. The LSTM network is employed for prediction tasks which leverages the ability to handle sequential IoT data streams and dropout layers are utilized to prevent overfitting. Finally, K-means clustering mechanism is used to establish an error threshold for anomaly detection. The performance of proposed LSTM and Sparse Autoencoder is evaluated on SKAB dataset which illustrated higher precision (0.824), recall (0.982) and f1 scores (0.847). The proposed method outperformed existing anomaly detection methods such as Neural Network (NN), Adaptable and Interpretable Framework for Anomaly Detection (AID).