Website picture tampering detection method based on deep learning

Abstract

The invention relates to a website picture tampering detection method based on deep learning. The method comprises the steps of collecting a sample picture with a text, establishing a training set after preprocessing, training a corresponding network model, further training a semantic similarity detection model and constructing a sensitive word bank, crawling and preprocessing a website picture tobe detected, obtaining a text content result in the website picture through the network model, calculating the semantic similarity between the text content result and the sensitive word bank throughthe semantic similarity detection model after word segmentation, and judging whether the sensitive content and the website picture are tampered based on the similarity. The text block recognition precision is improved; the method is accurate in text extraction, small in model, high in convergence speed and good in performance, semantic recognition is carried out on texts in pictures, sensitive speech in webpage pictures can be accurately detected, whether a website is invaded by hackers or not can be quickly and efficiently recognized, an alarm is given in time, the method plays an important role in maintaining website safety, and spreading of bad illegal information can be blocked in time.

Description

technical field [0001] The present invention relates to the technical field of general image data processing or generation, in particular to a method for detecting tampering of website pictures based on deep learning. Background technique [0002] With the accelerated development of the network, the importance of monitoring and supervising the network ecology is increasing. Hackers are always looking for opportunities to break into specific websites and perform a series of operations on them to achieve their criminal goals; what’s more, they are looking for website loopholes of important government enterprises, using methods such as implanting backdoors to control website servers, intruding Afterwards, it tampered with the page and published various illegal speeches by adding text content to the pictures. Based on the periodicity of its activities, it attacked two to three websites every week. After the webpage was tampered with, if it was not detected in time, the Governme...

AI Technical Summary

Problems solved by technology

[0004] For the detection method of single character extraction, the minimum operation unit is the character, and the single character is separated by segmentation or scoring the candidate area, and then the character is classified with the trained model. However, in more scenarios , the arrangement of characters on the picture is cumbersome and complicated, and it is extremely difficult to segment characters, and forced segmentation will destroy its physical structure. In addition, the segmentation of characters does not consider other characters before and after it, and the recognition accuracy in character recognition is greatly reduced.

[0005] For the detection method of text line extraction, compared with conventional objects, the aspect ratio of text lines varies greatly and there is an overall direction. The information extracted by the four-tuple description method of conventional object frame BBox is not sufficient, such as some WordArt uses curved text lines, and there are more handwritten fonts with more variations. These features make the effect of directly extracting text through text lines unsatisfactory.

Images