How to verify the authenticity of the P-passed certificate? Thirty million netizens posted photos with Messi? The popularization of image editing technology allows everyone to photoshop pictures, but it also brings difficulties in identifying "fake pictures" and even fraud.
To this end, Alibaba Security, together with the National Anti-Counterfeiting Engineering Center of Huazhong University of Science and Technology, and ICDAR, the only top conference in the direction of international document analysis and identification, held a screenshot tampering detection competition on the Tianchi platform, and opened up a special track of "finding the most powerful photo masters in the entire network". In this way, the fake image detection algorithm model is improved to improve the accuracy of fake image detection.
Zhou Yu, an associate professor at the School of Electronic Information and Communication of Huazhong University of Science and Technology, pointed out that certificate information and screenshot information are the main objects that are vulnerable to malicious tampering, posing a serious security threat to individuals and society. It is difficult to distinguish the authenticity of the falsified pictures, which presents a difficult and significant research task for the industry.
Caption: "Looking for the most powerful P-picture master on the whole network" special track part of the gameplay introduction
The pain points in life scenes are very important in this competition. To this end, Du Ming, a senior security expert at Ali, introduced: "The special track aims to lower the participation threshold of technical iterations, improve fun, and provide rich samples for professional tracks. Therefore, we divide the submitted competition pictures into screenshot areas, certificates District and other districts, hoping to be closer to the P-picture fraud scenarios that may be encountered in real life, so that the samples for training AI are more real, thereby improving its recognition ability."
Zhou Yu, an associate professor at the School of Electronic Information and Communication of Huazhong University of Science and Technology, believes that cooperating with enterprises to extract key issues from real applications and integrating them into competitions can help academia and industry with real tampered samples and promote Research and application of image counterfeiting technology.
For the general public who want to experience counterfeit identification technology, the competition opens the "Fake Image Crusher" public experience entrance: by uploading a picture, let AI identify whether the image has been tampered with, and present the results of identifying areas that may be tampered with in the form of heat maps.
"Fake Image Shredder" is a technology developed by researchers from Ali Security and the University of Macau. It will be launched in June 2022. It aims to improve netizens' awareness of network security and includes a number of image tampering detection technologies.
According to reports, image tampering detection technology conducts all-round analysis and statistics on images, and analyzes whether each area in the image has been compressed, resampled, introduced new feature pixels, etc. according to the characteristics of different image pixel areas, and marks different images through comparative analysis The difference points of the region, so as to find out the part of the image that has been tampered with.
The whole set of image tampering detection system will go through three steps of original image recognition, content tampering detection and content tampering location to complete the identification and judgment of whether the image is a tampered image and where it has been tampered with.
In the original image recognition stage, the image tampering detection system will make an original judgment on the file information of the image. Image modification and transmission often leave traces on the header file information, so the first step in the original image recognition needs to judge the originality of the image header file information.
For example, various time information tables are stored in the image header file. The time information in the original image is basically the same, but the time information of the image tampered with by software such as PS may appear contradictory, so it can be judged whether the image has been modified after shooting; The actual size of some non-original images may be inconsistent with the size recorded in the header file. Some heavy compression (usually after image modification will be dumped and recompressed) will change the actual width and height of the image, and sometimes the width and height recorded in the header file are not the same. No modifications were made, resulting in conflicting size information in the image header.
In the stage of content tampering detection, the image tampering detection system will detect whether the image content has been modified or not. For example, the compression characteristics and resampling traces of the image can be detected, the compression and storage history of the image can be traced, and the scaling factor of the image can be estimated.
For example, image modification or transfer will cause the JPG image to undergo secondary compression, and whether the image has been modified or transferred can be determined by detecting whether the image has secondary compression traces. Traces of secondary compression are difficult to find with the naked eye, and the DCT coefficients of secondary compression, due to the difference between the two quantization coefficients before and after, the histogram will show periodic changes. It is necessary to learn and classify the secondary compression by extracting the statistical properties of the DCT coefficients in the image.
Illustration: From left to right, the original image, the secondary compressed image, and the statistical characteristics of the secondary compressed DCT coefficients
In addition, the image tampering detection system can also perform statistical analysis on the edge consistency and content continuity of local objects in the image to determine whether the content in the image has been modified.
In the stage of content tampering location, through the judgment of the previous stage, the image tampering detection system combines the end-to-end AI algorithm to analyze the image content and features, and further marks which areas of the image have been modified.
Different from original image recognition, image content tampering detection and location need to have high robustness. In actual scenarios, there are often various operations for content tampering, including mosaic, region splicing, copy-paste, erasing, adding text, etc. There are also various types of images that need to be detected, including qualifications, certificates, software screenshots, product pictures, door face pictures and other images. In addition, image tampering localization also needs to be able to detect the content changes of the image after the image has undergone global post-processing. Common post-processing operations include cropping, scaling, recompression, blurring, filtering, recapping, and more. For different tampering operations, the means of detection are also different. The following is an example of an erasure type of tampering.
Taking erasing and tampering as an example, the erased area of the image is usually smoothed, so smooth feature extraction is performed on each area of the tampered image, and then the amplitude and gradient of pixel changes are analyzed to determine which areas of the image are tampered with. Common technical means include detection based on traditional image processing methods, such as difference, template matching, edge detection, etc., and methods based on deep learning, such as convolutional neural network (CNN). Among them, deep learning methods have attracted much attention due to their ability to automatically learn features and adapt to multiple tampering types.
Caption: Example of erasing text type tampering
Image tampering detection technology has important application value in today's digital age. Through the detection and analysis of the image content level, local object edge consistency and content continuity, and the combination of end-to-end AI algorithm for content tampering and location, it can effectively identify and locate whether the image has been tampered with, and ensure the authenticity and reliability of the image content. Reliability. With the continuous development of technology and the continuous expansion of application scenarios, image tampering detection technology will continue to be optimized and perfected to provide more accurate and reliable image information anti-counterfeiting and tampering detection and identification services for all walks of life.
Du Ming said that good technology must start from the real needs of society, focus on real problems, and solve real problems. With the help of such a competition, the participating samples and the excellent detection algorithm model generated by the competition are absorbed into the algorithm to improve the algorithm, which will help solve the problems faced by real social scenarios. He emphasized: "Building responsible and good technology is an important goal of Alibaba's technology ethics governance."
The competition address is attached: Fake Picture Crusher Challenge (aliyun.com)