A Lightweight CNN Activity Recognition And Attack Method

With the improvement of people's living standards,the demand for health monitoring and motion detection is increasing day by day.Therefore,it is of great significance to study human activity recognition methods.In the research of sensor-based human activity recognition,most of the studies use acceleration sensors to collect human activity data,and then combine the classification algorithm to complete the judgment of human activities.In the process of classifying human activities by extracting feature values in the original data,when extracting data features,it is easy to lose key information,thereby affecting the recognition accuracy.At the same time,the versatility of the existing recognition model is not strong,and the human body difference is not fully considered.The recognition algorithm is limited by the resources of the current handheld terminal or the wearable device,and is difficult to be transplanted into the handheld terminal or the wearable device.On the other hand,most researches on human activity recognition methods focus on the accuracy of recognition algorithms and the improvement of hardware device performance,ignoring the security problems inherent in them.Based on the above problems,this paper designs a lightweight active convolutional neural network(CNN)human activity recognition method,and studies a method of adversarial examples generation,which is beneficial to the follow-up activity recognition method to improve the defense ability against malicious attacks..The main work of the thesis includes the following two aspects:(1)Aiming at the problem that the traditional human activity recognition method loses part of the information in the process of extracting the feature value and affects the final recognition accuracy rate,a lightweight convolutional neural network human activity recognition method is studied,which will accelerate the original data.Converted to image format,the time series of acceleration data has a spatial relationship,and then use CNN feature extraction method to obtain key information,identify human activities,and finally compress the model using SVD to reduce the memory resources required by the classification model.Increase the speed of the operation,making it easier to port to portable devices.(2)Aiming at the problem of security protection of wearable devices during use,a method for generating adversarial examples is studied.The original data is randomly attacked and iteratively generated to produce the adversarial examples with the lowest accuracy of model identification.The human activity recognition method and the lightweight active convolutional neural network human activity recognition method proposed in this paper are used for identification and classification,and are used for anti-interference research of future human activity recognition algorithms.After experimental verification,the human activity recognition method proposed in this paper improves the recognition accuracy compared with the traditional human activity recognition method.At the same time,the adversarial examples attack method proposed in this paper has a good effect on various human activity recognition methods.