Research On Out-of-distribution Detection Based On Hyperellipsoid Decision Boundary

Out-of-distribution detection technology plays an important role in improving the security of the practical application of deep neural network model.Most of the existing deep neural network models are trained based on the closed-world assumption.The training samples and test samples used conform to the assumption of independent and identically distributed,which makes the deep neural network unable to recognize the out-of-distribution samples.Using the out-of-distribution detection technology to distinguish the in-distribution samples from the out-of-distribution samples is helpful to resist the attack risk of unknown samples by using the deep neural network model in practical applications such as autonomous driving.For the classification task of the nonlinear decision boundary,the deep multi-class data description describes the data distributed within the nonlinear decision boundary based on the hypersphere decision boundary,in which the data in the inner area of the closed hypersphere is detected as the in-distribution data,and the rest are detected as the out-of-distribution data.However,the out-of-distribution detection technology based on hypersphere decision boundary can not be applied to all distributions,because there are still a large number of irregular distributions in the real world,and the hypersphere decision boundary with equal radius everywhere can not fit these distributions compactly.Therefore,this thesis proposes a three-branch neural network model based on hyperellipsoid modeling and an algorithm named deep confidence geometry data description(DCGD).By using the decision boundary of the geometry description loss constraint model composed of confidence loss,distance loss and distribution disagreement loss,the spheres are tightened into ellipsoids,which is used to solve the problem that the application scenario of hypersphere model is single and can not fit the data distribution with irregular boundary shape compactly,and improves the detection ability of out-of-distribution samples and the classification ability of in-distribution samples.The experimental results show that the three-branch neural network model based on hyperellipsoid decision boundary and the deep confidence geometric data description algorithm can effectively detect out-of-distribution samples,and also achieve good performance in the task of classification of in-distribution samples.From the perspective of geometry,this thesis provides new research ideas and methods for the work in the field of out-of-distribution detection.The main innovations of this thesis are as follows:(1)A data description method based on hyperellipsoid decision boundary is proposed,and the confidence estimation matrix and confidence loss are given,so that the model can fit more complex data distribution.(2)A three-branch network structure is proposed to obtain the confidence loss,distance loss and distribution disagreement loss.It is used to calculate the geometric description loss to constrain and update the hyperellipsoid decision boundary.(3)A deep confidence geometry data description algorithm is proposed,which can improve the detection accuracy of the model without the participation of in-distribution samples in model training and verification.The parameter sensitivity analysis is carried out through ablation experiments.