Research On Image Encryption Algorithm Based On Dynamic Compressed Sensing

In the era of big data,there are many kinds of information carriers,and images are used in various fields,but the security problem comes with it.It is imperative to encrypt image information.Compressed sensing can reduce the amount of data and encrypt the image,so it has been deeply studied by researchers.Image encryption combined with compressed sensing can not only protect information security,but also reduce the overhead of image transmission in the network.At present,most of image encryption algorithms based on compressed sensing use chaotic system to construct random observation matrix.In order to improve the quality of compression and reconstruction,some algorithms also make sparse transformation on the image,but this does not break through the limitations of sampling rate and sparsity in compressed sensing,and the random observation matrix independent of image features is adopted,which makes it impossible to dynamically adapt image features during image compression and key information is lost.However,the traditional reconstruction algorithm performs poorly in reconstruction efficiency and quality.In addition,the scrambling diffusion encryption adopted by some algorithms after compression encryption has defects.Therefore,based on the theories of chaotic system,compressed sensing and convolutional neural network,the construction of observation matrix and image compression and encryption is studied in this thesis.The main contents are as follows:（1）Based on convolutional neural network,a compression network is designed as observation matrix for compression encryption.Based on the block compression theory,the image is evenly segmented before compression,and combined with the hash value of the sub block image,a two-dimensional Logistic chaotic random sequence is generated for scrambling and diffusion.Due to the periodicity of sliding scrambling,its step size is limited for encryption.Finally,the ciphertext is obtained by bidirectional XOR diffusion.Using nonlinear depth reconstruction network for reconstruction can effectively improve the quality of image reconstruction.The simulation results show that the proposed encryption algorithm has improved the encryption performance.The self-learning ability of neural network makes the algorithm no longer pay attention to image sparsity,and the sampling rate is not limited.It can dynamically and adaptively extract image features.The combination of compression network and reconstruction network also improves the quality of image reconstruction by 5%～10%.On the basis of（1）,an improved chaotic image encryption algorithm based on dynamic compressed sensing is proposed.Due to the block compression,the pixel dispersion of the sub block image is not high after limited period sliding scrambling encryption.The improved zigzag transform combined with sliding scrambling is used to get better scrambling effect.Based on the fact that three-dimensional logistic chaos still has the shortcomings of ergodicity and randomness,it is improved to improve the ergodicity and randomness of the pseudo-random sequence generated by it.The simulation results show that the histogram distribution of the improved chaotic encryption algorithm is more uniform,the information entropy is closer to the ideal value of 8,and the pixel correlation is closer to the ideal value of 0.At the same time,it has the ability to resist brute force attack,differential attack,noise attack and other damage.