Information steganography method in image animation process

Abstract

The invention provides an information steganography method in an image animation process, which comprises the following steps of: disguising a steganography tool as a deep neural network for executing an image animation task, converting the style of a given image into an animation style by utilizing a generator, and embedding secret data into the given image. And meanwhile, the embedded data can be extracted by utilizing the trained decoder. The generator, the decoder and the discriminator adopt an adversarial training mode, so that the generated image has an animation style, and secret information can be embedded and extracted. Due to the fact that secret information is embedded in the image animation process instead of being embedded through modified image pixels after training, the influence of information embedding on image animation is very small, and compared with a popular steganography algorithm, the secret information is more difficult to detect. In addition, the parameters of the decoders are pre-stored in the receiving end, so that network parameters do not need to be transmitted to the receiving end, and the secret information can be correctly extracted only by holding the corresponding decoders.

Description

Technical Field

[0001] The invention belongs to the technical field of communication encryption, and in particular relates to a method for steganography of information during image animation. Background Art

[0002] With the continuous development of informatization, information security issues have become increasingly important. Traditional means of ensuring information security are mainly achieved through cryptography. However, the use of cryptographic encryption methods cannot conceal the fact that communication is in progress. It is not applicable in specific situations such as copyright protection and covert communication. The information hiding technology developed in recent years has made up for this shortcoming of cryptography. Information hiding technology achieves the purpose of information security by embedding secret information in ordinary carriers. These carriers include common text, audio, pictures, and videos. Pictures have become the most commonly used carriers due to their widespread daily use. In the communication process of this cryptographic information, there are two main participants, the protector and the attacker of the information, corresponding to the steganalyst and the steganalyst. The purpose of the steganalyst is to embed the secret information into the carrier through the information hiding algorithm without being discovered by the enemy. It is usually necessary to make it visually, auditorily and statistically imperceptible; while the purpose of the steganalyst is to determine whether a specific carrier contains secret information. Generally speaking, once the steganalyst has reasonable suspicion of a secret carrier, the information hiding process is a failure. Summary of the invention

[0003] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a method for steganography of information during the image animation process.

[0004] The purpose of the present invention can be achieved by the following technical solutions:

[0005] The present invention provides a method for steganography of information during image animation, comprising the following steps:

[0006] Constructing a network architecture for information hiding, wherein the network architecture includes a generator, a decoder, and a discriminator;

[0007] Conduct adversarial training based on the loss functions of the generator, decoder, and discriminator to obtain the trained generator, decoder, and discriminator;

[0008] The sending end inputs the natural world image and the information to be hidden into the trained generator to obtain an animated stego image. The generator includes an animation module, an information acquisition module and a stego block. The natural world image and the information to be hidden are processed by the animation module and the information acquisition module respectively. The two outputs are connected in the channel dimension and then input into the stego block to output an animated stego image.

[0009] The sender transmits the animated stego-image through a public channel;

[0010] The receiving end obtains the animated stego image through the public channel, inputs the animated stego image into the trained decoder, and the decoder outputs the information to be hidden.

[0011] Furthermore, the generator and the decoder are pre-stored at the transmitting end and the receiving end.

[0012] Furthermore, the generator includes two side inputs, one side input is a natural world image, and the other side input is information to be hidden. The natural world image input side structure includes: two convolution blocks, a downsampling block, a convolution block, a depth separation block, a downsampling block, a convolution block, eight inverted residual blocks, a convolution block, an upsampling block, a depth separation block, a convolution block, an upsampling block, two convolution blocks, a standard convolution and Tamh activation block, and a mixed convolution block. The information to be hidden input side structure includes a steganalysis upsampling block, and the two side outputs are connected in the channel dimension and input into the steganalysis block to output an animated steganalysis image.

[0013] Furthermore, the steganographic block includes three sets of standard convolution plus LRelu activation plus instance normalization.

[0014] Furthermore, the convolution block includes standard convolution, instance normalization and LRelu activation, the upsampling block includes upsampling and depth separation blocks, the steganographic upsampling includes 3 groups of transposed convolution plus LRelu activation plus instance normalization, the mixed convolution block includes standard convolution, LRelu activation and instance normalization, the depth separation block includes depth-separable convolution, instance normalization, LRelu activation and convolution block, the inverted residual block includes convolution block, depth-separable convolution, instance normalization, LRelu activation and standard convolution, instance normalization, and the downsampling block includes downsampling and two depth separation blocks.

[0015] Furthermore, the decoder input is an animated stego image, and the output is the information to be hidden, and the decoder includes three sets of standard convolutions.

[0016] Furthermore, the discriminator input is an animated stego image or a real animated image, and the output is whether the image is an animated stego image. The discriminator includes standard convolution, LRelu activation, two sets of standard convolution plus LRelu activation plus standard convolution plus instance normalization plus LRelu activation, standard convolution, instance normalization, LRelu activation, and standard convolution.

[0017] Furthermore, the total loss function L(G,E) of the generator and decoder is:

[0018] L(G,E)=w adv E[(G(p i)-1) 2 ]+w con L con (G,D)+w gra L gra (G,D)+w col L col (G,D)+w dec L dec (G,E)

[0019] L con (G,D)=E[||VGG l (p i )-VGG l (G(p i ))||1]

[0020] L gra (G,D)=E[||Gram(VGG l (x i ))-Gram(VGG l (G(p i )))||1]

[0021] L col (G,D)=E[||Y(p i )-Y(G(p i ))||1+||U(p i )-Y(U(p i ))|| H +||V(p i )-V(U(p i ))|| H ]

[0022] Among them, G is the processing process of the generator, E is the processing process of the decoder, D is the processing process of the discriminator, and p i is a real-world image, x i is the grayscale image of the animation image, w adv 、w con 、w gra 、w col 、w dec They are 10, 1.5, 3, 30, 1, L con is the content loss, L gra (G, D) is the grayscale style loss, L col (G,D) is the color reconstruction loss.

[0023] Furthermore, the total loss function L(D) of the discriminator is:

[0024] L(D)=w adv [E[(D(a i)-1) 2 ]+E[(D(G(p i ))) 2 ]+E[(D(x i )) 2 ]+0.1E[(D(y i )) 2 ]]

[0025] Among them, a i is the animated image, y i It is the grayscale image of the smoothed animation image.

[0026] Furthermore, the adversarial training is performed according to the loss functions of the generator, decoder and discriminator. The specific training process includes: setting the learning rates of the generator and discriminator to 0.0002 and 0.0004 respectively, the number of training rounds to 70, the batch size to 6, using the Adam optimizer for optimization, using the real-world image as the content image, the content image and the secret information to be input into the generator to obtain the stego-image, and at the same time, inputting the stego-image and the real animation image into the discriminator for adversarial training.

[0027] Compared with the prior art, the present invention has the following advantages:

[0028] The generator, decoder and discriminator of the present invention adopt an adversarial training method, so that the generated image has an animation style and can embed and extract secret information. In addition, the present invention improves the undetectability of information by hiding information during the image animation process. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A module diagram used in the present invention;

[0030] Figure 2 It is a structural diagram of the steganographic model of the present invention;

[0031] Figure 3 It is a training process diagram of the model of the present invention;

[0032] Figure 4 It is a steganographic image effect diagram generated by the present invention;

[0033] Figure 5 It is a comparison chart of the undetectability of the present invention and other steganographic algorithms. DETAILED DESCRIPTION

[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0035] This embodiment provides a method for steganography of information during the image animation process, disguising the steganography tool as a deep neural network that performs the image animation task, using a generator to convert the style of a given image into an animation style and embed secret data into the given image, and at the same time, the trained decoder can be used to extract the embedded data. The generator, decoder and discriminator adopt an adversarial training method, so that the generated image has an animation style while embedding and extracting secret information. Among them, the generator and decoder are pre-stored at the sending end and the receiving end, the structure of the generator refers to Unet, the discriminator is used to distinguish whether the image is generated or a real animation image, and the decoder can restore the steganographic image to secret information. Because the network parameters of the generator, decoder and discriminator trained under different random initialization conditions are all different, the secret information of the steganographic image generated by a specific generator can only be extracted by a specific receiving end, ensuring the security of information hiding. Since the secret information is embedded in the image animation process instead of being embedded by modifying the image pixels after training, the information embedding has little effect on the image animation, and is more difficult to detect than the popular steganography algorithm. In addition, the parameters of the decoder are pre-stored at the receiving end, so there is no need to transmit the network parameters to the receiving end, and only the corresponding decoder is needed to correctly extract the secret information.

[0036] Secret information is embedded in the process of image animation to maintain the effect of image animation. In this embodiment, the generator, decoder and discriminator are obtained by adversarial training and pre-stored at the sending end and the receiving end. Therefore, there is no need to transmit the decoder to the receiver, and the secret information can be correctly extracted by holding the corresponding decoder. The generator, decoder and discriminator used in this embodiment are based on the following modules:

[0037] like Figure 1As shown, the convolution block includes standard convolution, instance normalization and LRelu activation, the upsampling block includes upsampling and depth separation blocks, the steganographic upsampling includes 3 groups of transposed convolution plus LRelu activation plus instance normalization, the mixed convolution block includes standard convolution, LRelu activation and instance normalization, the depth separation block includes depth separable convolution, instance normalization, LRelu activation and convolution block, the inverted residual block includes convolution block, depth separable convolution, instance normalization, LRelu activation and standard convolution, instance normalization, the downsampling block includes downsampling and two depth separation blocks, and the steganographic block includes three groups of standard convolution plus LRelu activation plus instance normalization. Wherein, "K" and "S" are the size and stride of the convolution kernel, respectively, "C", "H" and "W" are the number of channels, height and width of the feature map, respectively. "+" indicates the addition between elements of the feature map.

[0038] exist Figure 1 Based on this, the steganographic model structure of this embodiment is as follows:

[0039] use Figure 1 The network architecture for information hiding is constructed as follows: Figure 2 As shown in the figure, the structure of the generator G refers to Unet, the discriminator D is used to distinguish whether the image is generated or real anime image, and the decoder E can restore the steganographic image to the secret information. "×" indicates connection in the channel dimension.

[0040] Assume the generator process is G, the discriminator is D, the decoder is E, and the real-world image is p i , the animated image is a i , the grayscale image of the animation image is x i , the smoothed animation image is e i , further, e i The grayscale image is y i For content loss L con (G, D) and grayscale style loss L gra (G, D), use the pre-trained VGG19 as the perception network to extract high-level semantic features of the image. con (G,D) and L gra (G, D) can be expressed as equation (1) and equation (2). Convert the image color in RGB format to YUV format and construct the color reconstruction loss L col (G,D). col In (G, D), the Y channel uses L1 loss, and the U and V channels use Huber loss, which can be expressed as formula (3). dec (G,E) uses logarithmic cross entropy loss as the loss function of steganalysis embedding. The total loss function of the generator and decoder is formula (4), and the total loss function of the discriminator is formula (5). Where wadv , w con , w gra , w col , w dec are 10, 1.5, 3, 30, 1 respectively.

[0041] L con (G, D) = E[||VGG l (p i ) - VGG l (G(p i ))||1] (1)

[0042] L gra (F, D) = E[||Gram(VGG l (x i )) - Gram(VFF l (G(p i )))||1] (2)

[0043] L col (G, N) = E[||Y(p i ) - Y(G(p i ))||1 + ||U(p i ) - Y(U(p i ))|| H + ||V(p i ) - V(U(p i ))|| H (3)

[0044] L(G, E) = w adv E[(G(p i )) - 1) 2 + w con L con (G, D)+ w gra L gra (G, D)+ w col L col (G, D)+ w dec L dec (G, E) (4)

[0045] L(D) = w adv [E[(D(a i )) - 1) 2 + E[(D(G(p i ))) 2 + E[(D(x i )) 2 + 0.1E[(D(y i )) 2 (5)

[0046] In order to balance the training speed and training quality. The learning rates of the generator and discriminator are set to 0.0002 and 0.0004 respectively. The number of training rounds is 70, the batch size is set to 6, the Adam optimizer is used for optimization, and the Nvidia3090 GPU is selected for training. The training set uses 6656 real-world images as content images and 1792 animated images of the movie "The Wind Rises" directed by Hayao Miyazaki as style images. The resolution of all training images is set to 256×256. The construction process is as follows Figure 3 As shown, the content image and the secret information are input into the generator to obtain the stego-image. The secret information in the stego-image can be restored using the decoder. At the same time, the stego-image and the animation image are input into the discriminator for adversarial training, so that the stego-image has an animation style and contains the secret information.

[0047] like Figure 2 As shown in the figure, the present invention proposes a method for transmitting secret information during image animation. The training set uses 6656 real-world images as content images and 1792 animated images of the movie "The Wind Rises" directed by Hayao Miyazaki as style images. The resolution of all training images is set to 256×256. First, the embedding capacity of the secret information is calculated, and the steganographic images under various embedding amounts are demonstrated through experiments. Then the undetectability of the steganographic images is tested.

[0048] (1) Press Figure 2 The invention relates to a framework for transmitting secret information during image animation.

[0049] (2) The experimental environment is built with PyTorch and cuDNN, and the training is performed on a single Nvidia 3090 GPU running on a Windows 10 system. The learning rates of the generator and discriminator are set to 0.0002 and 0.0004, respectively. The number of training rounds is 70, the batch size is set to 6, and the Adam optimizer is used for optimization.

[0050] (3) The size of the secret information is m×32×32. The size of the stego-image is 256×256. The value of m is set to 4, 8, 16, 32, and 64, respectively. The corresponding embedding capacities are 4096, 8192, 16384, 32768, and 65536 bits, respectively. From equation (6), it can be calculated that the secret information bits per pixel are 0.0625, 0.125, 0.25, 0.5, and 1, respectively. Examples of the generated stego-images are as follows: Figure 4As shown, the “original image” represents the content image, the “animated image” represents the result of animation only without secret data, and the “steganographic image” represents the result of animation and embedding secret data at the same time; the corresponding embedding capacities of (a)-(e) are 4096, 8192, 16384, 32768 and 65536 bits, respectively.

[0051]

[0052] (4) For comparison, we use four popular steganographic algorithms, namely SUNIWARD (Spatially Generalized Wavelet Relative Distortion), WOW (Wavelet Obtained Weights), HILL (High Pass, Low Pass, Low Pass), and DFEI (Deformed Distortion Function for Enhanced Images), to embed the secret data into the image using only the embedding network for style transfer. The embedding capacity is set to t = [4096, 8192, 16384, 32768, 65536]. To test the undetectability of our method, we use four popular steganalysis schemes, namely SPAM (Subtracted Pixel Adjacency Matrix), SRMQ1 (Spatial Enrichment Model with Single Quantization Step), maxSRMd2 (Maximum Displacement Enrichment Model with Channel Awareness), and SCRMQ1 (Spatial Color Enrichment Model with Quantization Step). Half of the carrier image and the stego image are used for training, and the rest are used for testing. The evaluation criterion for undetectability is the minimum steganalysis error P obtained on the test set under the same prior conditions. E , defined in formula (7). FA and P MD Represent the false alarm rate and missed detection rate respectively. A higher P E This means higher undetectability. The undetectability results of the present invention and SUNIWARD, WOW, HILL and DFEI schemes are as follows: Figure 5 As shown, it can be seen that the steganographic algorithm of the present invention for the image after style transfer is superior to the undetectability of the existing steganographic algorithm.

[0053]

[0054] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program code.

[0055] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for steganography of information during image animation, characterized in that: The following steps are involved: Constructing a network architecture for information hiding, wherein the network architecture includes a generator, a decoder, and a discriminator; Conduct adversarial training based on the loss functions of the generator, decoder, and discriminator to obtain the trained generator, decoder, and discriminator; The sending end inputs the natural world image and the information to be hidden into the trained generator to obtain an animated stego image. The generator includes an animation module, an information acquisition module and a stego block. The natural world image and the information to be hidden are processed by the animation module and the information acquisition module respectively. The two outputs are connected in the channel dimension and then input into the stego block to output an animated stego image. The sender transmits the animated stego-image through a public channel; The receiving end obtains the animated stego image through the public channel, inputs the animated stego image into the trained decoder, and the decoder outputs the information to be hidden.

2. A method for steganography of information during image animation according to claim 1, characterized in that: The generator and the decoder are pre-stored at the transmitting end and the receiving end.

3. A method for steganography of information during image animation according to claim 1, characterized in that: The generator includes two side inputs, one side input is a natural world image, and the other side input is information to be hidden. The natural world image input side structure includes: two convolution blocks, a down sampling block, a convolution block, a depth separation block, a down sampling block, a convolution block, eight inverted residual blocks, a convolution block, an up sampling block, a depth separation block, a convolution block, an up sampling block, two convolution blocks, a standard convolution and Tamh activation block, and a mixed convolution block. The information to be hidden input side structure includes a steganalysis upsampling block, and the outputs on both sides are connected in the channel dimension and input into the steganalysis block, and an animated steganalysis image is output.

4. A method for steganography of information during image animation according to claim 3, characterized in that: The steganalysis block includes three sets of standard convolution plus LRelu activation plus instance normalization.

5. A method for steganography of information during image animation according to claim 3, characterized in that: The convolution block includes standard convolution, instance normalization and LRelu activation, the upsampling block includes upsampling and depth separation blocks, the steganographic upsampling includes 3 groups of transposed convolution plus LRelu activation plus instance normalization, the hybrid convolution block includes standard convolution, LRelu activation and instance normalization, the depth separation block includes depth-separable convolution, instance normalization, LRelu activation and convolution block, the inverted residual block includes convolution block, depth-separable convolution, instance normalization, LRelu activation and standard convolution, instance normalization, and the downsampling block includes downsampling and two depth separation blocks.

6. A method for steganography of information during image animation according to claim 1, characterized in that: The decoder input is an animated stego image, and the output is the information to be hidden. The decoder includes three sets of standard convolutions.

7. A method for steganography of information during image animation according to claim 1, characterized in that: The discriminator input is an animated stego image or a real animated image, and the output is whether the image is an animated stego image. The discriminator includes standard convolution, LRelu activation, two sets of standard convolution plus LRelu activation plus standard convolution plus instance normalization plus LRelu activation, standard convolution, instance normalization, LRelu activation, and standard convolution.

8. A method for steganography of information during image animation according to claim 1, characterized in that: The total loss function L(G,E) of the generator and decoder is: L(G,E)=w adv E[(G(p i )-1) 2 ]+w con L con (G,D)+w gra L gra (G,D)+w col L col (G,D)+w dec L dec (G,E) L con (G,D)=E[||VGG l (p i )-VGG l (G(p i ))||1] L gra (G,D)=E[||Gram(VGG l (x i ))-Gram(VGG l (G(p i )))||1] L col (G,D)=E[||Y(p i )-Y(G(p i ))||1+||U(p i )-Y(U(p i ))|| H + ||V(p i )-V(U(p i ))|| H ] Among them, G is the processing process of the generator, E is the processing process of the decoder, D is the processing process of the discriminator, and p i is a real-world image, x i is the grayscale image of the animation image, w adv 、w con 、w gra 、w col 、w dec They are 10, 1.5, 3, 30, 1, L con is the content loss, L gra (G, D) is the grayscale style loss, L col (G,D) is the color reconstruction loss.

9. A method for steganography of information during image animation according to claim 1, characterized in that: The total loss function L(D) of the discriminator is: L(D)=w adv [E[(D(a i )-1) 2 ]+E[(D(G(p i ))) 2 ]+E[(D(x i )) 2 ]+0.1E[(D(y i )) 2 ]] Among them, a i is the animated image, y i It is the grayscale image of the smoothed animation image.

10. A method for steganography of information during image animation according to claim 1, characterized in that: The adversarial training is performed according to the loss functions of the generator, decoder and discriminator. The specific training process includes: setting the learning rates of the generator and the discriminator to 0.0002 and 0.0004 respectively, the number of training rounds is 70, the batch size is set to 6, and the Adam optimizer is used for optimization. The real-world image is used as the content image, the content image and the secret information are input into the generator to obtain the stego-image, and the stego-image and the real animation image are input into the discriminator for adversarial training.

Images