Traffic sign adversarial sample detection method based on neighbor discrimination and classification device

Abstract

The invention discloses a traffic sign adversarial sample detection method based on neighbor discrimination and a classification device, and belongs to the field of machine learning. The method comprises: inputting a to-be-detected sample into a trained traffic sign classification network to obtain a first classification result and an intermediate feature map; inputting the intermediate feature map of the to-be-detected sample into a neighbor discrimination model, and taking a data set of the neighbor discrimination model as an intermediate feature map set of the training sample to obtain theprobability that the to-be-detected sample belongs to each traffic sign category; and judging whether the probability that the to-be-detected sample belongs to the first classification result is greater than or equal to a set threshold, if so, determining that the sample is a normal sample, otherwise, determining that the sample is an adversarial sample. AThe idea of calculating the similarity between the samples of the neighbor discriminant classification model and the simplicity and high efficiency of the method are utilized, on the premise of not influencing the classification performance of the deep neural network, the representation of an original data sample is changed, and the generation difficulty and attack difficulty of an adversarial sample are increased, so that the active defense capability of the system is improved, and the complexity of the model is not increased.

Description

technical field [0001] The invention belongs to the field of machine learning, and more specifically relates to a traffic sign adversarial sample detection method and a classification device based on neighbor discrimination. Background technique [0002] In the past few years, artificial intelligence scholars have made great breakthroughs, and deep learning has made cold machines "smart and sharp". Not only can they identify various types of objects from videos, but they can also identify pedestrians and traffic lights on the road in real life. Deep learning methods have made breakthroughs in image, voice, and natural language processing, and AI has surpassed humans in some fields. However, the researchers also found that systems based on deep neural network models can be easily fooled by adversarial examples. Adversarial samples are one of the biggest threats. Adversarial samples refer to samples that can cause machine learning models to make wrong judgments by adding inte...

AI Technical Summary

Problems solved by technology

However, these methods only make it more difficult to generate adversarial samples, and at the same time greatly increase the complexity of the original deep model, adding complex offline parameter update steps, and it is difficult to deploy in some fields with high real-time computing requirements.

Images