Generalization safety evaluation method for deep learning image classification model

Abstract

The invention discloses a generalization safety evaluation method for a deep learning image classification model, and belongs to the technical field of machine learning. At present, the important problem to be solved in the deep learning related research is to improve the robustness of the model while solving the security threat problem with generalization characteristics faced by the deep learning image classification model, the invention utilizes the generalization security evaluation method oriented to the deep learning image classification model, by testing indexes such as the active defense capability, the adversarial sample detection capability and the passive defense capability of the deep learning image classification model for adversarial samples, the security of the deep learningimage classification model is comprehensively evaluated, security holes existing in the model are explored in the evaluation process, and meanwhile, due to the generalization characteristic of the method, the method can be suitable for most deep learning image classification models, and the method has important theoretical and practical significance for improving the safety of the deep learning field.

Description

Technical field [0001] The present invention relates to a general safety assessment method for the depth learning image classification model, belonging to the technical field of machine learning. Background technique [0002] Deep Learning Technology is an important branch of machine learning (Machine Learning) technology, is an algorithm for characterizing learning of data in artificial neural network, also known as unsupervised feature learning. ), That is, there is no need for human design feature extraction, and the feature is learned from the data. Deep learning is essentially a nonlinear combination of multi-layer representation learning methods. In recent years, deep learning technologies have developed rapidly, emerged in a large depth learning framework, such as depth convolutional neural network (CNN), to generate anti-network (GaN), deep volume generation against network (DCGAN), etc., based on these frameworks Emergence, deep learning technology has been widely used i...

AI Technical Summary

Problems solved by technology

[0003] However, with the continuous expansion of the application scope of deep learning, its vulnerability in the face of adversarial examples needs to be resolved urgently.

Deep learning technology is generally vulnerable to adversarial samples. The disturbed sample input causes the model to give a wrong output with high confidence. In many cases, training on different subsets of the training set has different structures. The model will misclassify the same adversarial example, which means that the adversarial example becomes a blind spot of the training algorithm

Images