Research On Robustness Of Forgery Detection Method Based On Neural Network

With the rapid development of science and technology,image forgery techniques emerge in an endless stream.People from the very beginning,through manipulation,combined multiple negatives into a fake picture.Today,the average person can use some simple image editing software to edit pictures.Image forgery technology is becoming more and more advanced.Some malicious attackers use advanced forgery technology to forgery images.Such forgery images spread rapidly through social media platforms,causing people to misjudge objective facts,and even bringing many hidden security problems to the society and the country.Therefore,many advanced methods for image forgery detection have emerged.From the beginning of the traditional forgery detection methods to now based on neural network forgery detection methods.The detection performance of the forgery detection method is becoming more and more superior.According to previous studies,neural networks are vulnerable,so it is of practical significance to study the countermeasure robustness of neural network-based forgery detection method.On the other hand,the adversarial sample image generated by the adversarial sample generation technique is also considered to be a forged sample,thus stimulating us to establish a connection between the forged sample and the adversarial sample.There are two main tasks in image forgery detection based on neural network: first,forgery technology recognition.It is necessary to identify the counterfeit mode of the image content.Second,the location of the counterfeit area.You need to locate the fake area in the fake image.Although existing neural network-based forgery detection methods have superior identification performance on clean samples,there is a lack of research on the identification accuracy of disturbed samples.This paper considers that the shortcoming of the existing research on neural network-based forgery detection technology is the lack of research on the robustness evaluation of neural network-based forgery detection model.The objective of this paper is to study the robustness of neural network-based forgery detection methods.Firstly,a variety of neural network-based forgery detection methods are implemented.Secondly,the neural network forgery detection methods are attacked and experimental results are recorded.Specifically,the main work of this paper is as follows:(1)This paper realizes a variety of forgery detection methods based on neural network with excellent detection performance in recent years.A large number of experiments show that this type of forgery detection method has excellent detection performance on a variety of forgery detection data sets.(2)In this paper,different types of adversarial samples are generated for the data set used in the forgery detection method,and the generated adversarial sample test set can successfully deceive the neural network-based forgery detection model.A large number of experiments show that the neural network-based forgery detection model has a certain vulnerability.(3)This paper holds that adversarial sample images can also be considered as forged samples.The adversarial sample detection method is used to detect common forged samples to explore the relationship between forged samples and adversarial samples.Experiments show that adversarial sample detection method has poor effect on the detection of common forged samples.It may be because the internal feature distribution of adversarial samples is quite different from common forged samples,or it may be because the existing adversarial sample detection methods have certain limitations on the detection range of data samples,resulting in unsatisfactory detection effect.