Fund Project:Project supported by the National Natural Science Foundation of China (Grant Nos. 61475104, 61177009)
Received Date:08 December 2020
Accepted Date:24 February 2021
Available Online:12 July 2021
Published Online:20 July 2021
Abstract:Most of optical encryption systems are symmetric cryptosystems. The plaintext and the ciphertext in optical image encryption are related linearly. The security of the system needs to be strengthened. The asymmetric cryptosystem based on phase truncated Fourier transforms (PTFT) makes the security of the encryption system greatly improved by its nonlinear phase truncation. Deep learning (DL) as a method of machine learning was proposed decades ago. With the development of computer’s performance, the practicality of deep learning proves to be more and more obvious. Recently, deep learning has been effectively used in many fields such as biomedicine, object detection, etc. The good results have been achieved. In this article proposed is the attack to the PTFT encryption system by deep learning. Through the PTFT encryption system, we construct a plaintext-ciphertext paired image dataset and then train it by residual network (ResNet). There are two problems encountered by the traditional neural network model. One is vanishing or named exploding gradient, which makes training effect difficult to converge and the other is a degradation phenomenon. When continuing to increase the number of layers for a suitable depth model, the model accuracy will decline which is not caused by overfitting. This problem can be solved by the ResNet to a certain extent by directly bypassing and then taking the input information to the output to protect the integrity of the information. The biggest difference between ordinary directly connected convolutional neural networks and ResNet is that the ResNet has many bypass branches that directly connect the input to the subsequent layers, so that the subsequent layers can directly learn the residuals. The ResNet can automatically learn the decryption characteristics of the encryption system. Finally, the test set is used to test the decryption performance of the trained model. The data show that the model can restore the image with high quality and the model has a certain anti-noise ability. Compared with the two-step iterative amplitude recovery algorithm, the the method proposed in this paper can recover high quality image. Keywords:optical encryption/ phase truncated Fourier transforms/ deep learning/ residual network
此后, 针对密文含有$20\% $能量比高斯噪声的情况, 我们尝试让训练集中的密文也被与密文能量比为$20\% $的高斯噪声污染, 重新制作数据集并重新训练, 再用训练好的网络去处理含有$20\% $能量比高斯噪声的测试集中的密文, 结果如图6(b)所示. 图 6 使用含不同能量比高斯噪声的密文训练集后的测试效果 (a) 0%; (b) 20% Figure6. Test results after using ciphertext groups of Gaussian noise with different energy ratios: (a) 0%; (b) 20%.
作为对比, 图6(a)是训练集中密文没有被高斯噪声污染情况下训练好的网络, 对含有$20\% $能量比高斯噪声的测试集密文重建结果, 可以看到经过噪声训练的神经网络在抗噪声方面表现更佳. 因此在实际中, 如果待处理的密文含有噪声, 可以考虑通过让训练集中的密文也遭受大致相当的噪声污染的方法, 以提高网络的训练效果. 由于在实际应用中, 密文受到噪声污染的比例通常是无法确定的, 因此针对训练集和测试集噪声比例不同做出的测试如图7所示(测试集噪声比例均为30%). 图 7 使用含不同能量比高斯噪声的密文训练集后的测试效果 (a) 0%; (b) 20% Figure7. Test results after using ciphertext groups of Gaussian noise with different energy ratios: (a) 0%; (b) 20%.
可以看到, 即使训练集和测试集的噪声比例不同, 经过噪声训练的神经网络在恢复被噪声污染的密文时仍然效果更好. 为了评价图像质量, 我们使用峰值信噪比(peak signal to noise ratio, PSNR)和结构相似性(structural similarity index, SSIM)作为参考[18]: