Search

Power production is responsible for the introduction of eutrophication into surface water bodies via multiple routes. Therefore, water quality assessment is quite necessary for the resource protection of surface water bodies near power plants. At present there is an urgent need to develop automated methods for fast and accurate water quality assessment. In this paper three CNN-based classifiers are constructed aiming at the classification of surface water images. With respect to the threat of adversarial attacks, three typical attacking methods and a JPEG compression-based countermeasure are chosen for observing the impact of adversarial attacks. The evaluation of edge inference is made on two edge inference engines (platform-independent and platform-dependent) respectively. The experimental results show that the classifiers own excellent performance all with accuracy above 98%. They are able to be conveniently deployed on edge devices without performance degradation for at most 80 FPS. It is also shown that adversarial attacks will reduce classification accuracy markedly while JPEG compression can mitigate the threat as a simple and practical mechanism.

A smart grid may employ various machine learning models for intelligent tasks, such as load forecasting, fault diagnosis and demand response. However, the research on adversarial machine learning has attracted broad interest recently with the rapid advancement of deep learning techniques, which poses an evident threat to those deep learning models deployed in smart grids. In the face of the emergent problem, we make a compact survey of the adversarial attacks against deep learning models in smart grids. The research status of deep learning applications in smart grids and adversarial machine learning is briefly summarized firstly. Adversarial evasion and poisoning attacks in smart grids are analyzed and exemplified respectively with focus. To mitigate the threat typical countermeasures against adversarial attacks are also presented. From the survey it can be concluded that the threat of adversarial attacks in smart grids will be a kind of long-term existence and need continuous attention.

MOOC teaching has been developing rapidly in the context of COVID-19, and its teaching quality has become the focus of social attention. Therefore, this study analyzes the blended college English teaching model based on MOOC by constructing a teaching evaluation model. The co-occurrence rate index is used to improve the K-modes algorithm, and then the important index in the teaching mode is extracted. The neural network is used to construct the prediction model of student learning effect, which reflects the advantages and disadvantages of the teaching mode. Through experimental analysis, the accuracy of the improved K-modes algorithm in the model reaches 0.985. The recall rate reached 0.982; The average error of the prediction model is less than 1 in the error analysis. Therefore, the model accurately reflects the problems existing in the teaching model, and has a high prediction accuracy, indicating that the teaching evaluation model has a good evaluation effect. [ABSTRACT FROM AUTHOR]