Data enhancement-based out-of-distribution abnormal sample detection method

Abstract

The invention provides an out-of-distribution abnormal sample detection method based on data enhancement. A feature extraction stage: extracting feature vectors of samples in input distribution by using an encoder part of an automatic encoder algorithm; a data enhancement and reconstruction stage: a feature-based data enhancement method is used for transforming the features extracted in the feature extraction stage, and a decoder part of an automatic encoder is used for generating a sufficient number of auxiliary distribution external abnormal sample data sets containing semantic information from enhanced feature vectors; a sample marking stage: the prediction accuracy of an original classifier on a reconstructed sample is used as a soft label value of an abnormal sample outside distribution; and a classifier retraining stage: the classifier is retrained in combination with the in-distribution training data set containing the hard label supervision signals and the out-of-distribution abnormal sample data set containing the soft label supervision signals. According to the method, the uncertainty of the deep neural network model during prediction of the out-of-distribution abnormal samples is improved, and the safety of the model is improved.

Description

technical field [0001] The invention belongs to the field of credible machine learning, and in particular relates to a method for detecting out-of-distribution abnormal samples based on data enhancement. Background technique [0002] When machine learning classifiers are applied to real-world tasks, they tend to fail when the training and test dataset distributions are different. Worse, these classifiers often fail quietly while delivering highly confident predictions that are, unfortunately, incorrect. If classifiers can't point out when they might be wrong, it can limit their use and even lead to serious accidents. For example, a medical disease diagnosis model might consistently classify with high confidence, even though it should flag difficult examples that require human intervention. The resulting unlabeled misdiagnoses could hinder the future development of machine learning techniques in medicine. More generally and importantly, estimating when a model is wrong is ...

AI Technical Summary

Problems solved by technology

[0004] The purpose of the present invention is to provide a method for detecting out-of-distribution abnormal samples based on data enhancement to solve the technical problems of uncertainty in model prediction and low model security

Images