Research on Edge Intelligence-based Security Analysis Method for Power Operation System

A. Experimental setup
Our picture data set has about 100000 pictures, of which 80, 000 are for training and 20, 000 for testing. In our model, the vector dimension is 100, the convolution layer is 2 layers, the learning rate is 0.01, and the Adam optimization function is adopted. To compare our model with baselines, we use F1 as the evaluation metrics.
B. Experimental results at the edge
In the process of image recognition, we compare five latest methods, which are CNN, CNN+Pooling, CNN+Attention, Recurrent-CNN and Recurrent-CNN+Pooling, respectively. The experimental results are shown in Table I. We can see from the Table I that our method is better than the other five methods, and the results of p, R and F1 are all the best. In addition, the recognition effect of Recurrent-cNN is better than that of cNN, because Recurrent-cNN can better obtain the deep features of the image. Attention is better than pooling because Attention reinforces important information and weakens unimportant information.

3. 
   Cloud comprehensive evaluation experiment results
   

After the edge-side security detection, there will be multiple recognition results, these results are a time series, and there may be a relationship between them, so we use a graph neural network to process these results. Each node in the graph neural network is the previous recognition result. When the training is completed, these nodes are put into a regression function to obtain the regression result. We believe that these results are normal in a certain range, and the range of all results is counted to get the normal number. The ratio of these numbers to the total number is the stability of the system. However, if the result is less than a certain number, it means that it is abnormal. The ratio of these numbers to the total number is the failure rate of the system. The experimental results are shown in Table II. We can see from the table that the stability of our method is better than other methods, and the failure rate is lower than other methods.

4.