Abnormal Activity Detection in Edge Computing: A Transfer Learning Approach

Detecting real-time abnormal human activities require rapid and accurate computations. However, cloud computing can not meet the highly responsive requirement. Thus, the appearance of edge computing is promised to become a data processing solution for abnormal activity detection. On the other hand, artificial intelligence (AI), as a rapidly emerging force, is leading-edge computing to a new level. Edge computing embedded AI enables IoT devices to become smarter. In this work, we first propose a light deep learning framework using SMOTE to solve the imbalance label problem and implement a Convolutional Neural Network-Embedding-Feature (CNNEF) to recognize abnormal human activities through the sensor data in edge nodes. Then we feed the extracted high-level embedding features from CNNEF to the classical machine learning algorithms, such as logistic regression(LR), K-Nearest Neighbors (KNN), decision tree (DT), Naive Bayes (NB), Random Forests (RF) and support vector machines (SVM). Compare with previous research methods, and our comprehensive empirical results demonstrate the robustness of accuracy and notable time-saving of our algorithms. The highest AUC is 99.53%, and the best time reduction is 97.93%. The proposed method provides a new solution of edge computing-based AI for time-series IoT data applications.