Securing Image Security in the AI Era: Revealing Three Strategies to Solve the Fake Image Crisis

---

written in front

The 2023 World Artificial Intelligence Conference (WAIC 2023) will be held in Shanghai from July 6 to July 8, 2023. The theme of this conference is "Intelligently Connected World Creates the Future", focusing on the development of general artificial intelligence and encouraging the embrace of a new era of intelligence , to talk about the new future of the industry. They jointly discussed cutting-edge technologies in the field of artificial intelligence, including deep learning, machine learning, natural language processing, computer vision and other fields, and also demonstrated a number of application results and cases of artificial intelligence technologies in the real economy.

The conference includes a series of themed forums and seminars to discuss cutting-edge technologies and application scenarios in the field of artificial intelligence. In the theme forum of "Focus on the New Wave of AIGC in the Era of Large Models" held by the China Academy of Information and Communications Technology, Hehe Information showed us " ", 三大技术 一项标准helping to explore the multiple possibilities of the credible development of AI in the field of image content security:

From P-pictures to batch generation of fake pictures, AI image security has become the focus of credible AI

In recent years, with the rapid development of artificial intelligence (AI) technology, the application of AI in the image field has become increasingly widespread. But at the same time, some problems related to image security have also appeared, such as image tampering, false image generation, image steganography, etc. A large number of fraud cases and cyber violence incidents based on false images have caused adverse effects on a global scale.

By using deep learning models and large-scale datasets, a variety of highly realistic fake images such as fake news reports, fake scientific research, and fake advertisements can be easily generated. These false images often have a negative impact on people's thinking and behavior as well as society. False advertisements and false commercial promotions may deceive consumers and cause economic losses; false news may mislead the public's awareness of certain events. False social media identities or false brand images may lead to unpredictable consequences such as reduced consumer trust in brands or social media platforms. Now 图像内容的安全and 可信性have become the focus of public attention, and the security issues of AI images need to be resolved urgently.

Hehe Information has long been focusing on the cutting-edge technology exploration of "AI+DATA" in the field of document intelligence, as well as "fine-grained" visual difference forgery image identification, certificate document image information encryption, generative image identification, document image integrity protection and other industries The focus topic has outstanding advantages in helping individuals and enterprises protect the security of image content.

Three major technologies: advance layout, Hehe Information's AI image security technology helps the healthy development of the industry

Hehe Information's AI image security technology solution mainly includes three key technologies: AI image tampering detection, generative image identification, and OCR anti-attack technology, so as to deal with the increasingly frequent occurrence of malicious P pictures, generative fraud and illegal extraction of personal information, etc. .

✔ AI image tampering detection technology

The so-called image tampering detection is to give a picture, input it into the tampering detection model, and the model judges whether the picture has been tampered with, and locates the tampered area of ​​the tampered image. Hehe Information AI image tampering detection technology is an image tampering detection technology based on deep learning. This technology intelligently captures the subtle traces left by the image during the tampering process by learning the changes in statistical characteristics after the image has been tampered with, and displays the tampered location of the image area in the form of a heat map.

AI image tampering detection mainly uses two deep learning models: convolutional neural network (CNN) and recurrent neural network (RNN). Among them, CNN is used to extract the features of images, while RNN is used to analyze sequence data, such as time series information in text or images. In the process of image tampering detection, the input image is first preprocessed, including image size adjustment, feature extraction and data enhancement, etc.; then CNN is used to perform feature extraction on the image to obtain local features of the image, such as texture and structure; then Sequence analysis is performed on the extracted local features using RNN to capture the time series information of the image. In the model training phase, Hehe Information uses a supervised learning method to input known image tampering samples into the model, and let the model continuously adjust the weights and biases so that the output of the model is consistent with the real label (tampered or untampered ) as close as possible. Overall screenshot tampering detection is divided into four types:

• Copy move . Copy an area in an image to another area;
• stitching . Two unrelated images are stitched into a new image;
• erase . Mainly to erase some key information in the document;
• Reprint . Re-edit the new document based on the erasure.

This technology can detect various types of images, whether it is transfer records, transaction records, or screenshots of chat records, etc. Image tampering detection technology can recognize fakes with "intelligent eyes", which is very important for ensuring the authenticity and integrity of information and preventing fraud has a significant effect.

The specific operation of "reprinting" tampered image detection is as follows. Given a picture, input it into the combined information tamper detection model, and then it can determine whether the image has been tampered with, and locate the tampered area of ​​the tampered image.

The current main difficulty in tampering detection is whether the tampered screenshots can be found from pictures with no obvious differences, and whether the false positive rate in the found pictures is controllable; in fact, compared with certificate tampering detection, screenshot detection is more difficult. Existing vision models are usually difficult to fully explore the fine-grained difference features of the original image and the tampered image, so it is difficult to achieve a satisfactory accuracy rate. The image authenticity identification model based on the HRNet encoder-decoder structure proposed by Hehe Information combines the information of the image itself including but not limited to noise, spectrum, etc., so as to capture fine-grained visual differences to achieve high-precision identification, perfect solved these problems.

✔ Generative image identification technology

At the moment when artificial intelligence is booming, criminals are also beginning to use artificial intelligence to use generated pictures to evade copyright, identity verification, and illegally obtain benefits. Pictures generated by AI may cause confusion and misleading information, affecting people's judgment and cognition of real events and facts.

Combined information generative image identification technology is a technology that uses multi-dimensional features to distinguish real pictures from generative pictures. In practice, the main challenge faced by generative image identification technology is that there are many image scenes generated by AI, and it is difficult for machines to distinguish them. Hehe Information has developed a unique solution to this difficulty - based on spatial domain and frequency domain relationship modeling to identify AI-generated images.

This technology analyzes the characteristics and rules of the image, and uses the neural network to classify and identify the image to identify whether the image is generated by AI. It can identify many types of images, including but not limited to natural landscapes, portraits of people, icons, etc. By analyzing the details and structure of the image, the components generated by AI in the image are identified, and corresponding warnings or prompts are given. The following is the structure of the generative image discrimination model:

Based on spatial domain and frequency domain relationship modeling, multi-dimensional features can be used to distinguish subtle differences between real pictures and generated pictures without exhaustive pictures. After inputting a picture, the model focuses on spatial features through multiple spatial attention heads, and uses a texture enhancement module to amplify subtle artifacts in shallow features to enhance the model's perception and judgment accuracy of real and fake faces. Generative image identification technology can be perfectly applied to these industries:

1. Identity verification and access control : prevent the use of fake faces for identity verification, applied to safe passage systems, electronic access control systems, etc.;
2. Financial anti-fraud : Prevent fraudulent activities such as credit card fraud, account theft and identity fraud by using forged faces in the banking and financial fields;
3. Security detection of mobile devices : it can be used for the face unlocking function of mobile devices such as mobile phones to prevent forged faces from invading users' personal information and devices;
4. Digital image forensics : identify whether there are fake faces in images and videos, for forensic science and criminal investigation, etc.;
5. Remote Authentication for Video Conferencing : Ensure participants are authenticated and authenticated using their real faces.

✔ OCR anti-attack technology

In daily life, people often take pictures of their own relevant certificates and documents and send them to third parties, but the personal information carried on these pictures is likely to be identified, extracted and leaked by criminals using OCR technology, so a technology is needed to "Encrypt" the document image.

OCR (Optical Character Recognition, Optical Character Recognition) is a technology that converts printed or handwritten text into a computer editable and storage format. By using image processing and pattern recognition algorithms, the text characters in the image are extracted and converted into Converted to a numerical form that can be read by a computer. OCR adversarial attack is to give a test text picture and specify the target text, input it into the system for adversarial attack and output the result picture, so that the previously specified target text in the result picture cannot be recognized by the OCR system without affecting human eyes. Recognition of target text. The adversarial attack is divided into two forms : white-box attack and black-box attack , and black-box attack is usually the most commonly used.

As an image forgery detection technology based on deep learning, this technology uses neural network and deep learning methods to classify and identify images to identify the authenticity and credibility of images. Its main principle is to inject some well-designed fake features into the neural network, which are imperceptible to the human visual system, but can affect the classification results of the neural network. By adding these fake features to the real image, the original real image can be misjudged as a fake image by the neural network, so as to achieve the purpose of deception. It can prevent the machine from automatically crawling without affecting the naked eye. Through the application of OCR anti-attack technology, when we transmit pictures containing personal information on the Internet, the personal information in the image (such as address, bank card number, etc.) Carry out OCR against attacks to prevent pictures containing personal information from being intercepted by third-party platforms during network transmission, so as to protect personal privacy. Secondly, when the image contains a large amount of text data, OCR counterattacks can be performed on the data in the image to prevent third parties from reading and saving all the text content in the image through the OCR system, thereby reducing the risk of data leakage and playing a role in data encryption. Effect.

One standard: work with authoritative organizations such as China Academy of Information and Communications Technology to help science and technology for good

The China Academy of Information and Communications Technology has taken the lead in launching the formulation of the "Document Image Tampering Detection Standard", jointly compiled by Hehe Information, China Image and Graphics Society, University of Science and Technology of China and other scientific and technological innovation enterprises and well-known academic institutions. The purpose is to gather industry consensus around key issues such as forged image identification and provide effective guidance for the industry; to summarize the needs in the field of image content security, to explore the trend of document image tampering detection technology, and to help the healthy growth of the image industry.

The formulation of the "Document Image Tampering Detection Standard" may provide strong support for China's "trusted AI" system construction in the fields of machine vision and image processing. With the continuous development and application of artificial intelligence technology, we can expect more innovations and breakthroughs in the future, injecting more security and stability into the development of the industry.