Hypersphere Embedded Adversarial Training In Image Recognition

Obtaining reliable and robust neural networks has become a research hotspot.With the gradual popularization of the application of neural networks,image recognition systems begin to play an important role in various fields.With this comes the everchanging security threats in the complex environment.The attacks on models in key areas such as face recognition and signal recognition will bring great damage to the industry.Adversarial training is one of the most effective methods to improve model robustness,but its requirements for data,computing power and model size limit performance and application scenarios.On the basis of the research on the adversarial training,the research on the hypersphere embedded adversarial training based on feature angle is carried out in this paper.The main contributions of the project are as follows:(1)The concept of feature angle is proposed for the first time to solve many problems caused by the insufficient coding ability of the model,such as the large number of iterations adversarial samples,and the inability to accurately divide hard examples and simple examples.Feature angle refers to the angle between the decision boundary and the sample.The feature angle can quantify the index.By quantifying the angle between the decision boundary and the sample,the feature angle can reduce the number of sample iterations and distinguish between simple examples and hard examples.First,it is proved that the weight and decision boundary are orthogonal by hypothesis method,which provides a theoretical basis for the concept of feature angle.Then,on the basis of summarizing the existing interpretable methods,this paper proposes a new interpretable method the gap method between human and AI cognitive models.It explains that non robust features may still have key information.This shows that non robust features and robust features are equally important,and the cognitive gap between people and AI is the main reason for the lack of prior information encoding ability.We verified this by experiments on robust and non robust datasets.Therefore,the feature angle actually improves the model’s ability to encode prior information in solving problems such as more iterations,inability to accurately divide hard examples and simple examples.Finally,the feasibility and universality of feature angle are verified by experiments.The results show that the feature angle can be effectively coupled with the adversarial training,and can be applied to multiple data sets.(2)In view of many problems caused by simple example oriented training,such as imbalance between classes,insufficient class spacing and over fitting,we introduce the concept of characteristic included angle into the hypersphere embedded adversarial training.Through feature normalization and weight normalization,hard examples gradually dominate the training process.This method alleviates the imbalance between classes,increases the class spacing,and reduces the risk of over fitting.The visual saliency diagram and thermal diagram show that the hypersphere embedded adversarial training based on feature angle can make more effective use of non robust features than conventional adversarial training.Hard examples containing non robust features are more aggressive,and hard examples dominate training to further improve the coding ability of the model.We conducted experiments on three datasets.The results show that the hypersphere embedded adversarial training based on feature angle can actively adjust the learning dynamics,eliminate the accuracy loss caused by over fitting and other problems,and help to improve the model accuracy and convergence speed.