Adversarial sample-based verification code generation system and method

Abstract

The invention discloses an adversarial sample-based verification code generation system and method. The system comprises a verification code collection layer, a preprocessing layer, an adversarial sample generation layer, a sample splicing layer and a hybridization layer. The verification code acquisition layer transmits a collected verification code data set to the preprocessing layer, the preprocessing layer converts verification code pictures into a plurality of black and white character pictures through graying, binaryzation, denoising character segmentation and the like, and a constructedneural network and an adversarial sample algorithm are randomly selected from the adversarial sample generation layer to directionally generate adversarial samples. the sample splicing layer is usedfor splicing the single adversarial sample into verification codes with different lengths, and finally, the verification codes are reversedly preprocessed by the hybridization layer and are restored into colors, so that the verification codes aeCAPTCHA based on confrontation samples are generated. The system has the characteristics of low cost, low deployment difficulty and high attack resistance.The network attack can be more effectively resisted under the condition that the website does not need to replace an existing verification code system.

Description

technical field [0001] The present invention relates to the technical field of deep learning and information security, in particular to a verification code generation system and method based on confrontation samples. Background technique [0002] The full name of CAPTCHA is the FullyAutomaed Public Turing test to tell Computers and Humans Apart (which can be abbreviated as CAPTCHA), which is a public automatic program that distinguishes whether a user is a computer or a human. Its main function is to resist malicious robot programs, prevent spam comments in forums and blogs, filter spam, ensure the authenticity of online voting, and prevent malicious batch registration of websites. However, with the rise of convolutional neural networks in image recognition, the recognition rate of neural networks for simple graphic verification codes is basically close to 100%, and the recognition rate for complex verification codes also has a high recognition rate. Therefore, how to desig...

AI Technical Summary

Problems solved by technology

Although the second type reduces the recognition rate of the neural network to a certain extent, the user's pass rate is lower, which greatly affects the user experience

But why do the vast majority of websites use difficult-to-read complicated graphic verification codes and graphic verification codes without security? There are two main reasons: one is the cost, and the cost of replacing it with SMS verification codes and voice verification codes is relatively high; the other is that the difficulty factor is high. Compared with graphic verification codes, the new verification code technology system is complex to implement and requires high technical requirements.

Images