Research On Security Assessment Method Of Noise Image In Adversarial Environment

As an important transmission medium,images have been extensively used in daily life,commerce,and have been integrated into people’s life.With the continuous emergence of security issues related to privacy and national security,the problem of image security transmission has become prominent.Noise processing of images is one of the effective means to protect image security transmission.However,attackers are prone to launch direct attacks on noisy images in the adversarial environment;it is also possible to result in the wrong output of the model by adding subtle disturbance to the input example to attack the classification or recognition model.Therefor,the measurement of noise image security in adversarial environments is very important and has received widespread attention from researchers.Noisy images contain different levels of visual distortion from low to high,and image quality assessment algorithms are mainly used to evaluate high and medium visual quality distortion.Therefore,the existing image quality assessment algorithm cannot be directly used to measure the security of noisy images,and it is necessary to develop an evaluation algorithm specifically used to evaluate the security of noisy images.This thesis proposes a noise-based image security assessment algorithm by combining the human visual system(HVS)and deeply understanding the impact of noise on visual characteristics.According to whether the noise is easy to detect,it can be divided into security assessment of perceptible noise images and security assessment of imperceptible noise.The specific content includes the following two parts:(1)As one of the representatives of 3-D images,the light field image(LFI)has attracted wide attention because of their spatial-angle characteristics.With the continuous development of LFI research,its security has also been focused on.To detect the security of noise LFI,this thesis proposes a visual security measurement algorithm based on perceptible noise LFI.Due to the lack of an effective encrypted LFI dataset to evaluate the security of noise images,and the instability of training examples caused by a small amount of data is prevented,this thesis builds a perceptible noise LFI dataset.This contains 13 scenes,each scene contains 289sub-aperture images,a total of 90168 perceptible noise images.Based on this dataset,an image security assessment method for evaluating the security of noise image is proposed by combining the HVS,light field image and the advantages of depth learning in feature extraction.This algorithm obtains sub-aperture texture and structural features of LFI through deep learning to reflect changes in spatial information,and uses Gabor filtering to obtain epi-polar plane images to reflect changes in angular domain information.The experimental results demonstrate that compared to other classic and advanced evaluation algorithms,the method proposed in this thesis exhibits strong performance in evaluating the security of perceptible noisy images.(2)Unlike other 2-D images,adversarial examples are mainly generated by adding imperceptible noise to the input samples by attackers for attack classification or recognition models.To evaluate the security of adversarial examples,this thesis proposes an image security measurement algorithm based on imperceptible noise according to the characteristics of adversarial examples.Due to the lack of research on the measurement of the security of the adversarial examples and the lack of effective dataset to evaluate the security of the adversarial examples,this thesis constructs an imperceptible noise adversarial example.This dataset utilizes six countermeasures attack algorithms with five levels to generate 1050 adversarial examples.On the basis of this dataset,an image security evaluation algorithm based on a multi-scale feature extraction network is proposed by combining with the characteristics of the HVS.Experimental analysis shows that compared to mainstream image quality assessment algorithms,the method proposed in this thesis is closer to the perception of HVS.