Image encryption method and system based on bidirectional dynamic confusion and controllable noise diffusion

Abstract

The invention discloses an image encryption method and system based on bidirectional dynamic confusion and controllable noise diffusion, and belongs to the technical field of image encryption. The method comprises the following steps: acquiring a plaintext image, constructing and training a VortexCipheNet encryption network, and encrypting the plaintext image through the network based on a master key formed by a plurality of sub-keys to obtain a ciphertext image. Wherein the VortexCipheNet encryption network is formed by sequentially connecting a pixel grouping module, a first segmentation switching module, a noise injection module, a full connection layer, a first waveform activation function, a one-dimensional convolution layer, a second waveform activation function, a first-stage medium-range cavity convolution, a third waveform activation function, a second-stage medium-range cavity convolution, a Tanh activation function and a second segmentation switching module. Through the design of bidirectional dynamic confusion, controllable noise diffusion and a lightweight neural network, efficient, safe, dynamic and adaptive image encryption is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of image encryption, and in particular to an image encryption method and system based on bidirectional dynamic confusion and controllable noise diffusion. Background Art

[0002] With the rapid development of digital image technology, the storage and transmission of image data have become increasingly frequent. However, the privacy and security issues of image data have become increasingly prominent. Traditional image encryption methods mainly rely on classical encryption algorithms such as the Advanced Encryption Standard (AES), etc. Although these methods can ensure the security of data to a certain extent, with the improvement of computing power and the diversification of attack means, their security has been gradually challenged. In addition, traditional encryption methods usually require manual update and adjustment, and it is difficult to adapt to the dynamically changing security requirements. In recent years, the application of neural networks in the field of image processing has gradually attracted attention. Neural networks have powerful non-linear fitting ability and self-adaptive ability, which can provide new ideas for image encryption. However, most of the existing neural network-based image encryption methods have problems such as high model complexity and low encryption efficiency, and it is difficult to achieve dynamic adaptive encryption. Summary of the Invention

[0003] The present invention aims to provide an image encryption method and system based on bidirectional dynamic confusion and controllable noise diffusion, and realizes efficient, secure and dynamically adaptive image encryption through the design of bidirectional dynamic confusion, controllable noise diffusion and lightweight neural network.

[0004] The present invention is realized by the following technical solutions.

[0005] An image encryption method based on bidirectional dynamic confusion and controllable noise diffusion includes the following steps: Step 1: Obtain a plaintext image; Step 2: Construct and train a VortexCipherNet encryption network, where the VortexCipherNet encryption network is sequentially connected by a pixel grouping module, a first split-exchange module, a noise injection module, a fully connected layer, a first waveform activation function, a one-dimensional convolutional layer, a second waveform activation function, a first-level medium-range dilated convolution, a third waveform activation function, a second-level medium-range dilated convolution, a Tanh activation function, and a second split-exchange module; Step 3: Encrypt the plaintext image using the VortexCipherNet encryption network with the master key K. The master key K consists of the first sub-key K1, the second sub-key K2, the third sub-key K3, the fourth sub-key K4, and the fifth sub-key K5. The pixel grouping module groups the pixels of the plaintext image under the action of the first sub-key K1. The first splitting and swapping module splits the binary string under the action of the second sub-key K2. The noise injection module performs noise addition under the control of the third sub-key K3. The first waveform activation function, the second waveform activation function, and the third waveform activation function perform non-linear confusion operations on the feature sequence with the participation of the fourth sub-key K4 and the fifth sub-key K5; Step 4: Obtain the ciphertext image.

[0006] Further preferably, the process of encrypting the plaintext image using the VortexCipherNet encryption network is as follows: The pixel grouping module groups the pixels of the plaintext image under the action of the first sub-key K1, unfolds the grouped pixels bit by bit, and concatenates them with the first sub-key K1 to obtain an n-bit binary string; The first splitting and swapping module calculates the specific positions for splitting the n-bit binary string under the action of the second sub-key K2, performs splitting, and swaps the two parts of the sharded data; The noise injection module, under the control of the third sub-key K3, uses the third sub-key K3 as a seed to inject noise bound to the third sub-key K3 to generate a noise sequence of the same length as the input data; The fully connected layer globally mixes the data after splitting, swapping, and noise injection, maps the binary string after pixel grouping to the hidden space, and obtains the feature sequence; The waveform activation function performs non-linear confusion operations on the feature sequence with the participation of the fourth sub-key K4 and the fifth sub-key K5; The first-level medium-range dilated convolutional layer increases the receptive field of the convolutional kernel and enhances the transformation complexity in combination with the waveform activation function; The second-level medium-range dilated convolutional layer downsamples the feature sequence again through dilated convolution and non-linear waveform activation, and fuses different channel and position information; Finally, perform non-linear transformation on the feature sequence through the Tanh activation function to make the output features within the range of [-1, 1].

[0007] Further preferably, the fourth sub-key K4 and the fifth sub-key K5 are normalized according to the following formula: .

[0008] Further preferably, the specific steps for the first sub-key K1 to concatenate to obtain an n-bit binary string are as follows: The first step: Pixel grouping unfolds the grouped pixels into a binary string L1; Step 2: Connect it with the 32-bit first sub-key K1 to obtain a new binary string L2; Step 3: Adjust the binary string L2 to n bits in length.

[0009] Further preferably, the first waveform activation function, the second waveform activation function, and the third waveform activation function have the same structure, which is defined as: , where, is the output of the waveform activation function, both w1 and w2 are trainable activation parameters, and x is the input feature sequence.

[0010] The present invention also provides an image encryption system based on bidirectional dynamic confusion and controllable noise diffusion, including: A data acquisition module for obtaining a plaintext image; The VortexCipherNet encryption network is used to encrypt the input plaintext image. The VortexCipherNet encryption network is sequentially connected by a pixel grouping module, a first split-exchange module, a noise injection module, a fully connected layer, a first waveform activation function, a one-dimensional convolutional layer, a second waveform activation function, a first-level medium-range dilated convolution, a third waveform activation function, a second-level medium-range dilated convolution, a Tanh activation function, and a second split-exchange module. The pixel grouping module groups the pixels of the plaintext image under the action of the first sub-key K1. The first split-exchange module splits the binary string under the action of the second sub-key K2. The noise injection module performs noise addition under the control of the third sub-key K3. The first waveform activation function, the second waveform activation function, and the third waveform activation function perform non-linear confusion operations on the feature sequence with the participation of the fourth sub-key K4 and the fifth sub-key K5.

[0011] Advantages of the present invention: 1. Bidirectional dynamic confusion: Through pixel grouping and key binding (K1), the waveform activation function and the fourth key K4 and the fifth sub-key K5 achieve forward confusion. Through the split-exchange module and the medium-range dilated convolution, reverse confusion is performed. The forward confusion and the reverse confusion form a closed-loop confusion, and the attacker cannot restore the plaintext image through local analysis. By dynamically exchanging positions and bidirectional exchange of the input-output of the neural network, the anti-statistical analysis ability is enhanced.

[0012] 2. Controllable noise diffusion: Noise diffusion is performed under the control of the key, and the information entropy of the ciphertext image is close to the theoretical upper limit, enhancing the encryption effect.

[0013] 3. Lightweight neural network: The VortexCipherNet network has a total of seven layers, and the number of model parameters is 5K, which belongs to a lightweight neural network and is suitable for real-time encryption requirements.

[0014] 4. Dynamic adaptive encryption: Parameters such as the segmentation point, noise, waveform, etc. are all generated in real time by the key, and the encryption strategy is adjusted in real time through adversarial training to adapt to the dynamically changing security requirements.

[0015] 5. Coupling of the key and the algorithm: The encryption process is deeply bound to the key. Even if the algorithm structure is made public, different keys can produce completely different encryption behaviors, enhancing the anti-reverse engineering ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flowchart of the method of the present invention.

[0017] Figure 2 It is a diagram of the composition of the main key.

[0018] Figure 3 It is a network structure diagram of VortexCipherNet.

[0019] Figure 4 It is a flowchart of the operation of the segmentation and exchange module.

[0020] Figure 5 It is an example of a partial plaintext image.

[0021] Figure 6 It is an example of the encryption and decryption effect. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] The present invention will be further described in detail below with reference to the accompanying drawings.

[0023] As Figure 1 shown, this embodiment provides an image encryption method based on bidirectional dynamic confusion and controllable noise diffusion, and the steps are as follows: An image encryption method based on bidirectional dynamic confusion and controllable noise diffusion, and the steps are as follows: Step 1. Obtain the plaintext image; Step 2. Construct and train the VortexCipherNet encryption network; Step 3. Encrypt the plaintext image using the VortexCipherNet encryption network according to the main key K; Step 4. Obtain the ciphertext image.

[0024] In this embodiment, the main key K is composed of the first sub-key K1, the second sub-key K2, the third sub-key K3, the fourth sub-key K4, and the fifth sub-key K5, as Figure 2 shown.

[0025] As Figure 3As shown in the figure, the VortexCipherNet encryption network is composed of a pixel grouping module, a first segmentation and exchange module, a noise injection module, a fully connected layer, a first waveform activation function, a one-dimensional convolutional layer (Conv1D), a second waveform activation function, a first-level medium-range dilated convolution, a third waveform activation function, a second-level medium-range dilated convolution, a Tanh activation function, and a second segmentation and exchange module connected in sequence. The specific encryption process is as follows: The pixel grouping module groups the pixels of the plaintext image under the action of the first sub-key K1, unfolds the grouped pixels bit by bit, and connects with the first sub-key K1 to obtain a binary string with a length of n bits; The specific steps for connecting the first sub-key K1 to obtain a binary string with a length of n bits are as follows: The first step: Pixel grouping can be in groups of 2 pixels, or 4 or 8 pixels. And the grouped pixels are unfolded into a binary string L1; The second step: Connect with the 32-bit first sub-key K1 to obtain a new binary string L2. Its connection method is determined by the 1st and 2nd bits of the first sub-key K1, as shown in the following table:

[0026] The third step: Adjust the binary string L2 to a length of n bits. If L2 < n, perform the SHA256(K1) hash operation on the first sub-key K1, take the first n - L2 bits, and then splice the data into n bits. If L2 > n, truncate L2 to n bits; if L2 = n, directly output L2.

[0027] The first segmentation and exchange module calculates the specific position for splitting the n-bit binary string (this step directly performs a modulo operation on n) under the action of the second sub-key K2, performs segmentation, and exchanges the two parts of the sharded data; The noise injection module injects noise bound to the third sub-key K3 with the third sub-key K3 as the seed under the control of the third sub-key K3 to generate a noise sequence of the same length as the input data; The fully connected layer globally mixes the data after segmentation, exchange, and noise injection, and maps the binary string after pixel grouping to the hidden space; The waveform activation function performs a non-linear confusion operation on the data with the participation of the fourth sub-key K4 and the fifth sub-key K5; the first-level medium-range dilated convolutional layer increases the receptive field of the convolutional kernel and enhances the transformation complexity in combination with the waveform activation function; the fourth sub-key K4 and the fifth sub-key K5 are normalized according to the following formula: ; The second-level medium-range dilated convolutional layer downsamples the sequence again through dilated convolution and non-linear waveform activation, and fuses different channel and position information; Finally, the Tanh activation function is used to perform a non - linear transformation on the features; The pixel grouping module pre - processes the input data, converting the high - dimensional pixel data into a binary sequence that is more suitable for the processing of this network, so as to reduce the computational complexity of directly processing high - resolution images, thereby making the network load smaller. The pixel grouping module groups the pixels of the plaintext image under the action of the first sub - key K1, then unfolds the grouped pixels bit - by - bit, and then connects them with the first sub - key K1 to obtain a binary string with a length of n bits. The significant advantage of this design is that the influence of a single pixel bit can be diffused to multiple output bits through subsequent operations, achieving bit - level mixing. The participation of the first sub - key K1 makes the same input produce completely different binary representations under different keys. After grouping, adjacent pixels may be assigned to different groups, disrupting the original spatial neighborhood relationship and destroying the statistical characteristics of the plaintext image.

[0028] The first split - swap module, under the action of the second sub - key K2, calculates the specific position for splitting the n - bit binary string, then performs the split, and further exchanges the two parts of the sharded data. Its process is as Figure 4 shown. First, it checks the data shape of the input data. The data shape includes 1 - D (1 - dimensional), 2 - D (2 - dimensional), and 3 - D (3 - dimensional). Then, it operates according to the data shape. If the data shape is 1 - D, it performs data decompression operation; if the data shape is 2 - D, it directly processes; if the data shape is 3 - D, it performs data compression operation; obtains the split position, splits into the first shard and the second shard according to the split position, performs an exchange - connection operation on the first shard and the second shard, and then restores the data shape. After data compression operation (1 - D), maintaining the data shape (2 - D), and data decompression (3 - D) operations, the output is obtained. The split - swap module disrupts the data position, further disrupting the local correlation of the plaintext image and increasing the difficulty for the attacker to infer the original structure through statistical analysis. Further, it provides a more uniform input distribution for subsequent non - linear transformations, avoiding the leakage of local features. The split - swap module is used twice at the beginning and the end of the model to form a two - way position confusion, resisting position - based attacks.

[0029] The noise injection module injects noise under the control of the third sub - key K3. The randomized noise destroys the determinacy of the input data stream, making the same input produce different outputs during different runs; it hides the key patterns in the data, preventing the model from being reverse - engineered. Since the noise in the noise injection module is bound to the third sub - key K3, it can be regarded as a kind of dynamic key expansion, further increasing the randomness of the encryption process. Even if the attacker inputs the same plaintext image multiple times, due to the existence of noise, the output ciphertext images will be different. In addition, the noise is equivalent to an implicit regularization, preventing the model from overfitting. Therefore, this module also has a regularization effect.

[0030] In the fully connected layer, each input neuron is connected to all output neurons, and a weight matrix is used to linearly combine all dimensions of the input vector. A change in any single bit of the input will affect all output bits through the weight matrix, thus achieving the goal of cryptographic diffusion. By changing the dimensions, the fully connected layer projects the input data into a new space of different dimensions, further hiding the original statistical characteristics and achieving global confusion. After split exchange and noise injection, that is, after the input data undergoes split exchange and noise injection, the fully connected layer globally mixes it, maps the binary string after pixel grouping to the hidden space, and further disrupts the original bit distribution. Its confusion effect is further spread through the local receptive fields of the mid-range dilated convolutions in the subsequent two layers.

[0031] The waveform activation function is defined as: , where, is the output of the waveform activation function, w1 and w2 are both trainable activation parameters, and x is the input feature sequence.

[0032] The waveform activation function has non-linear confusion. The sine function and cosine function make the input and output highly non-linear, belonging to a many-to-one mapping. Therefore, an attacker cannot uniquely determine the input from the output. At the same time, the waveform activation function is calculated with the participation of the fourth sub-key K4 and the fifth sub-key K5. During encryption, the values of the fourth sub-key K4 and the fifth sub-key K5 determine the specific form of the waveform transformation. So the waveform function changes with different values of w1, w2, K4, and K5, and different keys will result in completely different confusion effects. Even if w1 and w2 are leaked and the attacker obtains some plaintext image-ciphertext image pairs, due to the randomness of K4 and K5, it is impossible to effectively infer the encryption results of other plaintext images, which also conforms to the Kerckhoffs design principle of cryptography.

[0033] The medium-range dilated convolutional layer increases the receptive field of the convolutional kernel through the dilation technique, enabling each output point to capture a wider input range. As a result, features can interact and correlate over longer distances, which is beneficial for increasing the complexity and non-linearity of cryptographic features. The use of the first-level medium-range dilated convolutional layer and the second-level medium-range dilated convolutional layer can increase the non-linear correlation of features at different levels, making it more difficult to analyze and predict cryptographic features. The dilation rate of the first-level medium-range dilated convolutional layer is 2, meaning that the actual covered input length is 5, i.e., the range length is 5, and the number of parameters is only the same as that of a 3×1 convolution. While maintaining computational efficiency, it also enhances cross-region information mixing. Additionally, from the perspective of the avalanche effect in cryptography, minor changes such as single-bit flips in the input will be propagated to multiple positions in the output through the range length of the dilated convolution. This jump-like propagation accelerates the avalanche effect, making the ciphertext image highly sensitive to changes in the plaintext image and the key. The first-level medium-range dilated convolutional layer further enhances the transformation complexity through a non-linear waveform activation function. The second-level medium-range dilated convolutional layer further enhances the feature complexity through dilated convolution and non-linear waveform activation. At the same time, the second-level medium-range dilated convolutional layer downsamples the sequence again, interweaving the inputs at adjacent positions into the same output point while compressing the number of channels to 1. This process fuses information from different channels and positions, increasing the unpredictability of the ciphertext image. This multi-level design is equivalent to multiple rounds of iteration, increasing the feature complexity and making it more difficult to break cryptographic features in cryptography. The design of the two-level medium-range dilated convolutional layer can also extract features at different scales, enabling cryptographic features to be effectively represented at different scales, thereby increasing the multi-scale nature of cryptographic features and further improving the security of the cipher. Finally, a non-linear transformation of the features is performed through the Tanh activation function, making the output features within the range of [-1, 1], increasing the boundedness and normalization of the features.

[0034] Another embodiment of the present invention provides an image encryption system based on bidirectional dynamic confusion and controllable noise diffusion, including a data acquisition module and a VortexCipherNet encryption network. The data acquisition module is used to obtain the plaintext image, and the VortexCipherNet encryption network encrypts the input plaintext image.

[0035] Experiments were conducted using the BOSSBase dataset, which consists of 10,240 grayscale images obtained by 7 digital cameras. Figure 5 Six of the plaintext images are given. Figure 6 For Figure 5 the encryption / decryption effect of the plaintext image A, Figure 6Among them, a is the plaintext image, b is the ciphertext image, c is the decrypted image, d is the horizontal correlation coefficient of the ciphertext image, e is the vertical correlation coefficient of the ciphertext image, and f is the diagonal correlation coefficient of the ciphertext image. Figure 5 The peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of the encryption / decryption of the 6 plaintext images are shown in Table 1, and the entropy of the plaintext image and the ciphertext image is shown in Table 2.

[0036] Table 1 PSNR / SSIM of Decrypted Images

[0037] Table 2 Entropy of Plaintext Images and Ciphertext Images

[0038] It can be seen from the experimental results that the present invention has achieved a very good encryption effect. The average PSNR of the decrypted image obtained by the decrypting party is 62.113 dB, and the average SSIM is 0.9889. The average entropy of the ciphertext image is 7.9902, which is close to the theoretical value of 8, and the average entropy of the plaintext image is increased by 7.71 percentage points.

[0039] The key length is the sum of the lengths of the four sub-keys, that is, 32 + 32 + 32 + 32 = 128, and the key space is as high as 2 128 , meeting the mainstream security requirements.

[0040] The above only expresses the preferred embodiments of the present invention, and does not limit the present invention in other forms. Any person skilled in the art may use the disclosed content to make changes or modifications into equivalent embodiments with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution content of the present invention still belong to the protection scope of the technical solution of the present invention.

Claims

1. An image encryption method based on bidirectional dynamic confusion and controllable noise diffusion, characterized in that It includes the following steps: Step 1: Obtain the plaintext image; Step 2: Construct and train the VortexCipherNet encryption network, which is sequentially connected by a pixel grouping module, a first split-exchange module, a noise injection module, a fully connected layer, a first waveform activation function, a one-dimensional convolutional layer, a second waveform activation function, a first-level medium-range dilated convolution, a third waveform activation function, a second-level medium-range dilated convolution, a Tanh activation function, and a second split-exchange module; Step 3: Encrypt the plaintext image using the VortexCipherNet encryption network according to the main key K, where the main key K is composed of a first sub-key K1, a second sub-key K2, a third sub-key K3, a fourth sub-key K4, and a fifth sub-key K5. The pixel grouping module groups the pixels of the plaintext image under the action of the first sub-key K1, the first split-exchange module splits the binary string under the action of the second sub-key K2, the noise injection module performs noise addition under the control of the third sub-key K3, and the first waveform activation function, the second waveform activation function, and the third waveform activation function perform non-linear confusion operations on the feature sequence with the participation of the fourth sub-key K4 and the fifth sub-key K5; Step 4: Obtain the ciphertext image.

2. The image encryption method according to claim 1, wherein The process of encrypting the plaintext image using the VortexCipherNet encryption network is as follows: The pixel grouping module groups the pixels of the plaintext image under the action of the first sub-key K1, expands the grouped pixels bit by bit, and connects with the first sub-key K1 to obtain an n-bit-long binary string; The first split-exchange module calculates the specific positions for splitting the n-bit-long binary string under the action of the second sub-key K2, performs splitting and exchanges the two parts of the sharded data; The noise injection module, under the control of the third sub-key K3, uses the third sub-key K3 as a seed to inject noise bound to the third sub-key K3 to generate a noise sequence of the same length as the input data; The fully connected layer globally mixes the data after split-exchange and noise injection, maps the binary string after pixel grouping to the hidden space, and obtains the feature sequence; The waveform activation function performs non-linear confusion operations on the feature sequence with the participation of the fourth sub-key K4 and the fifth sub-key K5; the first-level medium-range dilated convolutional layer increases the receptive field of the convolutional kernel and enhances the transformation complexity in combination with the waveform activation function; The second-level medium-range dilated convolutional layer downsamples the feature sequence again through dilated convolution and non-linear waveform activation, and fuses different channel and position information; Finally, the Tanh activation function performs non-linear transformation on the feature sequence to make the output feature within the range of [-1, 1].

3. The image encryption method according to claim 2, wherein The fourth sub-key K4 and the fifth sub-key K5 are normalized according to the following formula: .

4. The image encryption method according to claim 2, wherein The specific steps for the first sub-key K1 to connect to obtain an n-bit-long binary string are as follows: The first step: The pixel grouping expands the grouped pixels into a binary string L1; The second step: Connect with the 32-bit-long first sub-key K1 to obtain a new binary string L2; The third step: Adjust the binary string L2 to n bits long.

5. The image encryption method according to claim 3, wherein The first waveform activation function, the second waveform activation function, and the third waveform activation function have the same structure and are defined as: ， Among them, is the output of the waveform activation function, both w1 and w2 are trainable activation parameters, and x is the input feature sequence.

6. An image encryption system based on bidirectional dynamic confusion and controllable noise diffusion, characterized in that, It includes: Data acquisition module, used to obtain plaintext images; VortexCipherNet encryption network, used to encrypt the input plaintext images. The VortexCipherNet encryption network is composed of a pixel grouping module, a first split-exchange module, a noise injection module, a fully connected layer, a first waveform activation function, a one-dimensional convolutional layer, a second waveform activation function, a first-level medium-range dilated convolution, a third waveform activation function, a second-level medium-range dilated convolution, a Tanh activation function, and a second split-exchange module connected in sequence. The pixel grouping module groups the pixels of the plaintext image under the action of the first sub-key K1. The first split-exchange module splits the binary string under the action of the second sub-key K2. The noise injection module performs noise addition processing under the control of the third sub-key K3. The first waveform activation function, the second waveform activation function, and the third waveform activation function perform non-linear confusion operations on the feature sequence with the participation of the fourth sub-key K4 and the fifth sub-key K5.

7. The image encryption system according to claim 6, wherein The process of encrypting plaintext images using the VortexCipherNet encryption network is as follows: The pixel grouping module groups the pixels of the plaintext image under the action of the first sub-key K1, unfolds the grouped pixels bit by bit, and connects them with the first sub-key K1 to obtain an n-bit long binary string; The first split-exchange module calculates the specific positions for splitting the n-bit long binary string under the action of the second sub-key K2, performs splitting, and exchanges the two parts of the sliced data; The noise injection module, under the control of the third sub-key K3, uses the third sub-key K3 as a seed to inject noise bound to the third sub-key K3 to generate a noise sequence of the same length as the input data; The fully connected layer globally mixes the data after split-exchange and noise injection, maps the binary string after pixel grouping to the hidden space, and obtains a feature sequence; The waveform activation function performs non-linear confusion operations on the feature sequence with the participation of the fourth sub-key K4 and the fifth sub-key K5; the first-level medium-range dilated convolutional layer increases the receptive field of the convolutional kernel and enhances the transformation complexity in combination with the waveform activation function; The second-level medium-range dilated convolutional layer downsamples the feature sequence again through dilated convolution and non-linear waveform activation, and fuses different channel and position information; Finally, the Tanh activation function performs non-linear transformation on the feature sequence to make the output features within the range of [-1, 1].

Images