Research On Quantification Method Of CNN Predictive Uncertainty Based On Evidence Theory

With the rapid development of artificial intelligence,deep learning technology continues to mature.Convolutional neural network(CNN)models are increasingly used in safety-critical areas,and the reliability of model predictions is extremely important for high-safety and high-risk systems.In the application of safety-critical systems,it is not only required that the model be able to make predictions with high confidence,but also need to clearly indicate the reliability of the prediction,that is,to express the uncertainty of the prediction,so as to prevent catastrophic consequences when the system completely relies on the confidence of the model prediction to make decisions.Therefore,studying the quantification method of CNN predictive uncertainty is of great significance to the reliability guarantee of the safety-critical areas.Evidence theory,as a mathematical method to deal with uncertain reasoning problems,has the ability to express "uncertain" and "unknown".It can model and reason uncertain information without prior information.There are significant advantages in the presentation and modeling of information.This thesis has carried out the research on the quantification method of CNN predictive uncertainty based on evidence theory,and applied it to the detection task of adversarial samples and out of distribution samples.The main research contents of the thesis are:Firstly,this thesis studies the construction method of the evidence classifier.The core idea is to map the feature vector of the input sample to the evidence weight,and then reasonably transform the evidence weight into the basic probability assignment function and make decision.Based on this,this thesis combines the evidence theory with the CNN model,and carries out the reconstruction of the evidence classifier.Based on the output information of the evidence classifier,a quantitative evaluation method for the uncertainty of CNN prediction is proposed.Conflict and uncertainty are used to quantify the uncertainty of CNN prediction from two different angles.This method is simple to implement and has the advantages of strong scalability.Experiments prove that the method proposed in this thesis can effectively quantify the uncertainty of CNN prediction.Secondly,this thesis proposes an adversarial sample detection method based on the uncertainty quantitative evaluation method.By considering the characteristics of adversarial samples,this thesis designs different detection schemes and detection strategies.The three detection schemes are used to detect the adversarial samples with conflict value,uncertainty and combination scheme.The detection strategy includes global threshold and sub-category determination threshold.The former considers the overall situation of the sample,the latter considers the differences between samples of different categories.Experiments show that the AUC of this method on the Lenet-5adversarial samples generated by FGSM can reach 93.7% on the MNIST dataset.The detection strategy that determines the sub-category threshold value is better than the detection strategy of the global threshold value.Finally,this thesis proposes an out of distribution sample detection method based on the uncertainty quantitative evaluation method.According to the characteristics of the out of distribution sample and the principle of quartiles,this thesis chooses different detection thresholds and adopts different detection schemes for out of distribution samples.Detection.Experiments have proved that this method can effectively detect out of distribution.On the mixed sample composed of the MNIST data set and the Fashion MNIST data set,the detection AUC can reach 93.9%.