foreword
In today's society, images are one of the important ways of information dissemination and expression. However, with the advancement of technology, people can easily use various image editing software to tamper and forge pictures, and create scenes that seem real but are actually false.
This phenomenon has undoubtedly brought a series of negative effects to society. First of all, a large number of fraud cases based on false pictures emerge in an endless stream. Through fake pictures disguised as real scenes, criminals can deceive victims into obtaining their personal information or property, causing serious financial losses. Second, incidents of cyber violence also occur frequently. People can use image editing software to create false events, spread rumors, and maliciously target specific individuals or groups of people without actually happening . This has led to social tension and the proliferation of negativity, posing a threat to people's mental health and social stability.
The best solution is to defeat magic with magic (to defeat technology with technology) . Therefore, it is necessary for us to understand the "black technology" of image tampering detection that Hehe Information launched for us at the World Artificial Intelligence Conference.
Article Directory
AI Image Tampering Detection Technology
In our daily life, we all share some screenshots of chat records, some red envelopes with relatives and friends, and transfer money, but has this information been tampered with? Based on this information, Hehe Information has specially carried out thematic research on image documents. The research topic direction includes four categories, copying and moving (information in one picture is copied to another place); stitching (two irrelevant pictures are stitched together); erasure (erasing some key information in the image document); reprinting (editing some new documents on the basis of erasure).
This technology of Hehe Information is an image tampering detection technology and related systems based on deep learning. By learning the changes in the statistical characteristics of the image after it has been tampered with, it can intelligently capture the subtle traces left by the image during the tampering process, and display the tampered location of the image area in the form of a heat map. The related technology has been applied in banking, insurance and other fields.
The following is a demonstration of the application effect of Hehe Information's image tampering detection technology:
"Reprinting" tampered image detection: Input a picture into the AI image tampering detection model of Hehe Information, and the model can judge whether the image has been tampered with, and locate the tampered area of the image. The effect is shown:
Hehe Information also proposes an image authenticity identification model based on the encoder-decoder structure of HRNet, which combines the information of the image itself, including but not limited to noise, spectrum, etc., so as to capture fine-grained visual differences and achieve high-precision identification.
AIGC generates images causing problems
In the preface, it also mentions the negative problems that false images bring to society, and popularizes here what AIGC is. AIGC (Artificial Intelligence Generated Content) is image content generated by artificial intelligence. In recent years, with the rapid development of technologies such as deep learning and generative adversarial networks (GANs), AIGC has been widely used in various fields.
So, how does it generate the image?
First, AIGC's process of generating images is usually based on a neural network model. It uses a large-scale dataset to train the model, giving it the ability to learn image features and styles . Then, by tuning the model's parameters and input noise, new images with realistic appearance can be generated.
In addition, the images generated by AIGC can have various styles and contents, such as human faces, natural scenery, animals, etc. These images are so visually similar to real images that it is often difficult to tell whether they are real photos or generated by AI.
The most important point is that AIGC generates images in a wide range of applications . In terms of artistic creation, artists can use AIGC-generated images as a source of inspiration to break through traditional creative restrictions; in the design and advertising industry, AIGC-generated images can be used for product rendering, advertising, etc.; in the field of game development, AIGC-generated images can be used for scene and character design; in virtual reality and augmented reality technologies, AIGC-generated images can provide users with a richer and more realistic virtual experience.
Therefore, the production and misuse of these false images may lead to misleading information, violation of our privacy, and many other risks.
Generative Image Discrimination
As mentioned above, criminals may use the generated pictures to evade copyright and identity verification, illegally obtain benefits, and cause property losses to the people.
Based on this, for some situations where the similarity between the generated image and the real image is extremely high, and the problem that the generated image scene is too numerous to be exhaustive, Hehe Information considers the relevance of the spatial domain and the frequency domain in the algorithm, and comprehensively evaluates the multi-dimensional features. This comprehensive evaluation can more accurately distinguish small differences between generative images and real images, thereby improving the realism and accuracy of discrimination.
The model structure is shown in the figure below:
When we input a picture, the model focuses on the spatial features through multiple spatial attention heads, and uses the texture enhancement module to amplify the subtle artifacts in the shallow features to enhance the model's perception and judgment accuracy of real faces and fake faces.
Whether in finance, media, or other industries, generative image detection technology can reduce certain infringement problems and reduce the probability of fund theft. It has played a major role in anti-fraud, copyright protection and other fields.
OCR against attack technology
What is OCR?
OCR, the English full name is Optical Character Recognition, optical character recognition, it is a technology that converts printed or handwritten text into editable text. It recognizes and converts characters in images into computer-readable text by using image processing and pattern recognition algorithms.
OCR technology has a wide range of applications in extracting personal information from ID documents. By inputting the ID card image into the OCR system, the personal information on the ID card can be automatically identified and extracted, including name, gender, ethnicity, date of birth, ID number, etc.
It mainly includes several steps:
-
Image preprocessing: Preprocessing the ID card image, including image enhancement, cropping, denoising, etc., to improve image quality and character edge definition.
-
Area positioning: through image processing algorithms, determine the location of the area where the personal information on the ID card is located. These fields usually include name, gender, ethnicity, date of birth, ID number, etc.
-
Character segmentation: After determining the personal information area, each character needs to be segmented from the image for subsequent character recognition.
-
Character recognition: For each character, use OCR algorithm to convert it into readable text. Common OCR algorithms include methods based on template matching, methods based on deep learning, and so on. These algorithms compare characters to pre-trained models to determine the recognition result for each character.
-
Post-processing: There may be errors or inconsistencies in the recognized characters. In order to improve the recognition accuracy, post-processing operations can be performed. For example, use a verification algorithm to verify the validity of the ID number, or combine dictionaries and language models for error correction.
Therefore, based on the confidentiality requirements of personal and corporate business documents, Hehe Information has carried out innovative technology exploration and developed OCR anti-attack technology to "encrypt" document pictures.
The following figure shows the application effect of Hehe Information OCR anti-attack technology, which can cover key information and prevent machines from automatically crawling without affecting the naked eye:
OCR anti-attack technology can disturb the text of the scene or the text in the document without affecting the viewing and judgment of the naked eye, and "attack" the content containing key information such as Chinese, English, numbers, etc., prevent third parties from reading and saving all the text content in the image through the OCR system, reduce the risk of data leakage, and achieve the purpose of protecting information.
Promote the establishment of industry development standards for image content security, and help the systemic implementation of trusted AI
China Academy of Information and Communications Technology took the lead in launching the formulation of the "Document Image Tampering Detection Standard", aiming to establish the development order of the image content security industry and provide support for the systematic implementation of trusted AI. Hehe Information is one of the important participants in the standard .
The standard will focus on issues such as fine-grained visual difference forged image identification, generative image identification, and document image integrity protection, to gather industry consensus and provide effective guidance for the industry.
As a company deeply involved in the field of intelligent text recognition and intelligent image processing, Hehe Information has been recognized in intelligent document processing, and obtained the highest rating through the assessment of China Academy of Information and Communications Technology.
In the future, Hehe Information will work with academic and industrial partners to promote the safe and credible development of the AI image content industry .
The article ends here first, and we will continue to share relevant knowledge points in the future.
