〇. Preface
On the afternoon of July 7, the 2023 World Artificial Intelligence Conference (WAIC) "Focus·New Wave of AIGC in the Era of Large Models—Trusted AI" forum was held in the Red Hall of Shanghai World Expo Center. Famous experts, scholars, investors and leading entrepreneurs in the frontier fields of artificial intelligence and other technologies gathered together to explore the driving force of China's technological innovation.
At this World Artificial Intelligence Conference, China Academy of Information and Communications Technology held a special forum on the theme of "Trusted AI with multi-modal basic large models" . AIGC security has become an important topic in the current research and development of artificial intelligence.
Article Directory
1. Trustworthy artificial intelligence
Trusted AI refers to the concept of emphasizing credibility and reliability in the application of artificial intelligence.
He Jifeng, academician of the Chinese Academy of Sciences, believes that "artificial intelligence can no longer be regarded as simple technical work, but as the engine of change for the entire organization. Trustworthy artificial intelligence must be committed to ensuring data security and credibility, system behavior can be held accountable, and algorithm models can be trusted." Interpretation, credibility of the network environment, credibility of legal ethics and other issues.”
With the continuous development of artificial intelligence technology, more and more people have put forward requirements for the transparency, fairness and controllability of artificial intelligence systems. Trusted artificial intelligence needs to meet the following requirements:
-
Transparency: The decision-making process and algorithms of the artificial intelligence system should be explainable and understandable, and users should be able to understand how the system makes judgments and decisions.
-
Fairness: AI systems should avoid discriminatory treatment of different groups or individuals, and should not make unfair decisions based on personal attributes such as race, gender, and age.
-
Privacy protection: When processing personal data, AI systems should protect the privacy of users and ensure that personal information is not misused or leaked.
-
Security: Artificial intelligence systems should have the ability to prevent malicious attacks and abuses to ensure the security and reliability of the system.
-
Controllability: AI systems should have certain controllability, and users should be able to supervise and manage the system to ensure that the system behavior meets user expectations and meets legal and ethical requirements.
2. AI Image Security
Professor Xie Hongtao, a professor of the Department of Electronic Engineering and Information Science, University of Science and Technology of China, once mentioned, "In the era of AIGC outbreaks, ordinary people can use AI algorithms to generate high-quality text, images, and audio content. Among them, Midjourney, Stable Diffusion The realism of the images produced by such image generation methods is amazing, and it is difficult for human eyes to distinguish between real and fake. This can not help but arouse people's hidden worries: a large number of false pictures will be widely spread on the Internet. The proliferation of false pictures will cause a variety of social security problems”
The behavior of using AI to falsify images is a key issue that needs to be addressed in the development of AI credibility.
At present, the deep learning and algorithms of artificial intelligence are becoming more and more advanced, and the application of deep synthesis is growing exponentially. The behavior of abusing this technology is bringing ethical issues visible to the naked eye. In addition to using AI for face-changing fraud, some people also use it to make indecent videos of others, pretending to be celebrities to bring goods, etc. Today, with increasingly rich payment scenarios, the use of biological information such as human faces, voices, fingerprints, and irises is becoming more and more widely used, and risks such as privacy leaks are also increasing.
Just like that Internet buzzword: You never know whether the opposite of the Internet is a person or a dog. If such technologies are not restricted, AI technology will not only easily threaten personal and property safety, harm personal dignity and privacy, but may also pose greater social harm.
Image is an important medium for carrying information, and it is also an important focus of information security. With the development and popularization of image editing software, the threshold for making false images is gradually lowered, resulting in a large number of fraud cases and cyber violence incidents based on false images that have had a serious impact on a global scale.
With the rise of generative counterfeiting technology, people are paying more and more attention to the safety and credibility of image content. However, in the field of image content, "trusted AI" is just getting started.
At the forum, Hehe Information's AI image content security technology solution has received widespread attention from the industry. This solution can accurately locate traces of tampering in screenshots, intelligently distinguish generative images, and prevent criminals from using technical means to illegally crawl information on images.
Hehe Information has been committed to exploring cutting-edge technologies in the field of document intelligence, paying special attention to the application of "AI+OCR" technology. The identification technology of visual difference forged images is studied, and the solution to encrypt the image information of certificate documents is devoted to providing. In addition, they also pay attention to industry-focused issues such as generative image discrimination and document image integrity protection. Help individuals and enterprises enhance the security of image content through technical output and product services.
At the Trusted AI Forum of the World Artificial Intelligence Conference, Hehe Information demonstrated "three major technologies, one standard" and explored multiple possibilities for the development of trustworthy AI in the field of image content security.
Hehe Information's AI image security technology solution mainly includes three key technologies: AI图像篡改检测, 生成式图像鉴别, OCR对抗攻击技术, to deal with the increasingly frequent phenomenon of malicious P-pictures, generative fraud and illegal extraction of personal information.
1️⃣AI image tampering detection technology
Tang Qi, general manager of the Intelligent Innovation Department of Hehe Information, introduced that at the World Artificial Intelligence Conference last year, Hehe Information demonstrated the "PS tampering detection" technology for the first time, which can identify PS traces in various important commercial materials such as documents and bills. With "eyes of fire".
Based on deep learning, the technology provides image tampering detection technology and related systems. This technology intelligently captures the tiny traces left by the image during the tampering process by learning the statistical feature changes after the image has been tampered with. Use the form of heat map to display the tampered area in the image. This technology has been successfully applied in banking, insurance and other fields.
This year, the image tampering detection technology continued to be optimized and upgraded, and expanded to the field of "screenshot tampering detection". In the past, image tampering detection technology was mainly applied to natural scene images, but in fact, tampered images such as qualification certificates, documents, and screenshots are more likely to pose risks to people's lives. Hehe Information's artificial intelligence image tampering detection technology can detect a variety of screenshots, such as transfer records, transaction records, chat records, etc. Image tampering detection technology can be accurate whether it is through "copy and move" image tampering means of "picking out" key elements from the original image and moving them elsewhere, or using "wiping", "reprinting", etc. Identify fake images.
The main difficulty of this technology is whether it is possible to find tampered screenshots from pictures without obvious visual differences, and whether the false detection rate in the found pictures is controllable. Compared with certificate tampering detection, screenshot detection is more difficult. Therefore, Hehe Information proposes a model based on the encoder-decoder structure of HRNet. The model combines the information of the image itself, including but not limited to noise, spectrum, etc., to capture fine-grained visual differences, and fully utilizes the advantages of HRNet to achieve high-precision image authenticity identification.
2️⃣ Generative Image Identification
In recent years, artificial intelligence-generated content represented by language generation and visual generation has triggered social discussions on the credibility of graphic content. Criminals use the generated pictures to evade copyright, identity verification and other means, illegally obtain benefits and bring property losses to people.
In this context, Hehe Information has developed AI-generated image identification technology, which aims to help individuals and organizations identify and judge whether generated images are generated by AI, prevent false virtual person fraud, and solve the problems faced by generative AI. Ethical issues to promote the healthy development of generative AI.
There are two main difficulties in this project: one is that the generated image scenes are very diverse, which cannot be solved through exhaustive methods, nor can they be trained one by one through specific field segmentation; the second is that some generated images are similar to real pictures are so high that it is difficult to distinguish them. By establishing a relationship model between the spatial domain and the frequency domain, Hehe Information can use multi-dimensional features to distinguish small differences between real pictures and generated pictures without exhaustive images.
The model structure is shown in the figure below: 
the model uses multiple spatial attention heads to focus on the spatial features of the input image, and uses a texture enhancement module to highlight subtle artifacts in shallow features, thereby improving the accuracy of real and fake faces. Perception and judgment accuracy.
Generative image detection technology has a wide range of applications in anti-fraud, copyright protection and other fields. For example, in the financial industry, lawbreakers may use artificial intelligence synthesis technology to steal online funds and threaten citizens' property safety. This technology can reduce the risk of fund theft by intervening in the payment process. In the media industry, some picture suppliers use software to automatically generate pictures such as posters, deliberately conceal the source and sell them to third parties, and the third parties commercially use these pictures without knowing it, causing copyright infringement. Related detection techniques can solve these problems to a certain extent.
3️⃣OCR anti-attack technology
For the needs of life and work, people will take pictures of their own relevant certificates and documents and send them to third parties. The personal information carried on these pictures may be identified, extracted and leaked by criminals using OCR technology.
In order to meet the needs of individuals and enterprises for the confidentiality of documents, Hehe Information has explored innovative technologies and developed OCR anti-attack technology for "encrypting" document images. In order to prevent the above situation from happening, OCR anti-attack technology is used to interfere with the scene text or the text in the document, and "lock" the important files of individuals or enterprises, thus preventing third parties from reading and saving the text in the picture through the OCR system. All text content reduces the risk of data leakage and realizes the protection of private information.
This technology can disturb the scene text 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, and numbers. Doing so can prevent third parties from reading and saving all text content in the image through the OCR system, thereby reducing the risk of data leakage and protecting information security. At the same time, this technology will not affect the viewing and judgment of human eyes, so that the readability of the document will not be affected.
3. Summary
The development of artificial intelligence faces both positive and negative possibilities, and the battle between criminals and guardians is going on every day invisibly. The key lies in which side has stronger technical strength, which determines whether the Dao is one foot tall or the devil is one foot tall.
The realization of the new technology of AI image security benefits from the foundation accumulated by Hehe Information in the field of document images for more than ten years. As an artificial intelligence company, its strong scientific and technological strength is the confidence for Hehe Information to break through industry problems and cultivate innovative technologies. It is this scientific and technological strength that enables Hehe Information to maintain a competitive edge in this ever-changing field and provide society with safer image solutions.
At this stage, Hehe Information has obtained more than 100 invention patents with independent intellectual property rights, and provides intelligent text recognition products and AI services for hundreds of millions of users in more than 200 countries and regions around the world. In the future, Hehe Information will continue to pay attention to the new needs and trends in the image field, and use "excellent" AI to inject a sense of security into the development of the industry.
