A method for steganography of information during image animation

Abstract

This invention proposes a method for steganography during image animation. The steganography tool disguises itself as a deep neural network performing the image animation task. A generator converts the style of a given image to an animated style and embeds secret data into the given image. Simultaneously, a trained decoder extracts the embedded data. The generator, decoder, and discriminator employ adversarial training, ensuring that the generated image possesses an animated style while simultaneously embedding and extracting secret information. Because the secret information is embedded during image animation rather than through post-training pixel modification, the impact of information embedding on image animation is minimal, making it more difficult to detect compared to popular steganography algorithms. Furthermore, the decoder parameters are pre-stored at the receiver, eliminating the need to transmit network parameters to the receiver; simply possessing the corresponding decoder is sufficient to correctly extract the secret information.

Description

Technical Field

[0001] This invention belongs to the field of communication encryption technology, specifically relating to a method for steganography of information during image animation. Background Technology

[0002] With the continuous development of information technology, information security has become increasingly important. Traditional methods of ensuring information security mainly rely on cryptography. However, cryptographic encryption methods cannot conceal the fact that communication is taking place, making them unsuitable for specific situations such as copyright protection and covert communication. Information hiding technology, which has developed in recent years, effectively complements this shortcoming of cryptography. Information hiding technology achieves information security by embedding secret information into ordinary carriers, including common text, audio, images, and video, with images being the most commonly used carrier due to their widespread daily use. In this process of communicating encrypted information, there are two main participants: the information protector and the attacker, corresponding to the stegwriter and the steganalyst. The stegwriter's goal is to embed secret information into a carrier using information hiding algorithms without being detected by the enemy, typically achieving visual, auditory, and statistical imperceptibility. The steganalyst's goal is to determine whether a specific carrier contains secret information. Generally speaking, if a steganalyst has reasonable suspicion about a carrier containing secret information, the information hiding process has failed. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method for steganography in the process of image animation.

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

[0005] This invention provides a method for steganography during image animation, comprising the following steps:

[0006] Construct a network architecture for information hiding, the network architecture including a generator, a decoder and a discriminator;

[0007] Adversarial training is performed based on the loss functions of the generator, decoder, and discriminator to obtain the trained generator, decoder, and discriminator.

[0008] The sending end inputs a natural world image and the information to be hidden into a trained generator to obtain an animated steganographic image. The generator includes an animation module, an information acquisition module, and a steganographic 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 concatenated in the channel dimension and then input into the steganographic block to output an animated steganographic image.

[0009] The sending end transmits the animated steganographic image through a public channel;

[0010] The receiver acquires animated steganographic images through a public channel, inputs the animated steganographic images into the trained decoder, and the decoder outputs the information to be hidden.

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

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

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

[0014] Furthermore, the convolutional block includes standard convolution, instance normalization, and LRelu activation; the upsampling block includes upsampling and depth separation blocks; the steganography upsampling includes three sets of transposed convolutions plus LRelu activation plus instance normalization; the hybrid convolutional block includes standard convolution, LRelu activation, and instance normalization; the depth separation block includes depth-separable convolution, instance normalization, LRelu activation, and convolutional blocks; the inverted residual block includes convolutional blocks, depth-separable convolution, instance normalization, LRelu activation, standard convolution, and instance normalization; and the downsampling block includes downsampling and two depth separation blocks.

[0015] Furthermore, the decoder takes an animated steganographic image as input and outputs the information to be hidden. The decoder includes three sets of standard convolutions.

[0016] Furthermore, the discriminator takes either an animated steganographic image or a real animated image as input and outputs whether the image is an animated steganographic 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] Where G represents the generator's processing, E represents the decoder's processing, D represents the discriminator's processing, and p... i For real-world images, x i For the grayscale image of the animation, w adv w con w gra w col w dec The values ​​are 10, 1.5, 3, 30, and 1 L respectively. con For content loss, L gra (G,D) represents the grayscale style loss, and L... col (G,D) represents 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 For animated images, y i This is a smoothed grayscale image of the animation.

[0026] Furthermore, the adversarial training based on the loss functions of the generator, decoder, and discriminator specifically 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 real-world images as content images, inputting the content images and secret information into the generator to obtain steganographic images, and simultaneously inputting the steganographic images and real animation images 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 this invention employ adversarial training, enabling the generated images to have an animated style while embedding and extracting secret information. Furthermore, this invention improves the undetectability of information by hiding information during the image animation process. Attached Figure Description

[0029] Figure 1 This is a block diagram used in this invention;

[0030] Figure 2 This is a structural diagram of the steganography model of this invention;

[0031] Figure 3 This is a diagram illustrating the training process of the model of this invention;

[0032] Figure 4 This is a diagram showing the effect of the steganography generated by this invention;

[0033] Figure 5 This is a comparison chart of the undetectability of this invention with other steganography algorithms. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0035] This embodiment provides a method for steganography during image animation. The steganography tool is disguised as a deep neural network performing the image animation task. A generator converts the style of a given image to an animated style and embeds secret data into the given image. Simultaneously, a trained decoder extracts the embedded data. The generator, decoder, and discriminator are trained adversarially, enabling the generated image to have an animated style while embedding and extracting secret information. The generator and decoder are pre-stored at the sending and receiving ends. The generator's structure references Unet, the discriminator distinguishes between generated and real anime images, and the decoder restores the steganographic image to its 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 a 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 during image animation rather than after training by modifying image pixels, the impact of information embedding on image animation is minimal, making it more difficult to detect compared to popular steganography algorithms. Furthermore, the decoder's parameters are pre-stored at the receiving end, so there is no need to transmit network parameters to the receiving end; simply possessing the corresponding decoder is sufficient to correctly extract the secret information.

[0036] Secret information is embedded during image animation to maintain the animation effect. In this embodiment, the generator, decoder, and discriminator are obtained through adversarial training and pre-stored at the sending and receiving ends. Therefore, it is unnecessary to transmit the decoder to the receiver; simply possessing the corresponding decoder is sufficient to correctly extract the secret information. The generator, decoder, and discriminator used in this embodiment are based on the following modules:

[0037] like Figure 1As shown, the convolutional block includes standard convolution, instance normalization, and LReLU activation; the upsampling block includes upsampling and depth separation blocks; the steganography upsampling includes three sets of transposed convolutions plus LReLU activation and instance normalization; the hybrid convolutional block includes standard convolution, LReLU activation, and instance normalization; the depth separation block includes depth-separable convolution, instance normalization, LReLU activation, and a convolutional block; the inverted residual block includes a convolutional block, depth-separable convolution, instance normalization, LReLU activation, standard convolution, and instance normalization; the downsampling block includes downsampling and two depth separation blocks; and the steganography block includes three sets of standard convolutions plus LReLU activation and instance normalization. Wherein, "K" and "S" represent the kernel size and stride, respectively, and "C", "H", and "W" represent the number of channels, height, and width of the feature map, respectively. "+" indicates element-wise addition of the feature map.

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

[0039] use Figure 1 The construction of a network architecture for information hiding, such as Figure 2 As shown, the generator G is structured based on Unet, the discriminator D distinguishes whether an image is generated or a real anime image, and the decoder E can restore the steganographic image to its secret information. "×" indicates a connection along the channel dimension.

[0040] Assume the generator's processing 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 is x i The smoothed animation image is e i Furthermore, e i The grayscale image is y i Regarding content loss L con (G,D) and grayscale style loss L gra (G,D), using a pre-trained VGG19 network as the perceptual network to extract high-level semantic features of the image. L con (G,D) and L gra (G,D) can be expressed as equations (1) and (2). The image colors in RGB format are converted to YUV format, and the color reconstruction loss L is constructed. col (G,D). In L col In (G,D), the Y channel uses L1 loss, and the U and V channels use Huber loss, which can be expressed as equation (3). dec (G,E) uses log-cross-entropy loss as the loss function for steganographic embedding. The total loss function for the generator and decoder is given by equation (4), and the total loss function for the discriminator is given by equation (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] To balance training speed and quality, the learning rates for the generator and discriminator were set to 0.0002 and 0.0004, respectively. The number of training epochs was 70, the batch size was set to 6, and the Adam optimizer was used for optimization. An Nvidia 3090 GPU was selected for training. The training set used 6656 real-world images as content images and 1792 animation images from Hayao Miyazaki's film *The Wind Rises* as style images. All training images were set to a resolution of 256×256. The construction process is as follows... Figure 3 As shown, the content image and secret information are input into the generator to obtain the stegana image. The secret information in the stegana image can be restored using the decoder. At the same time, the stegana image and the animated image are input into the discriminator for adversarial training, so that the stegana image has an animated style and contains secret information.

[0047] like Figure 2 As shown, this 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 from Hayao Miyazaki's film *The Wind Rises* as style images. All training images are set to a resolution of 256×256. First, the embedding capacity of the secret information is calculated, and experiments are conducted to demonstrate steganographic images with various embedding capacities. Then, the undetectability of the steganographic images is tested.

[0048] (1) Press Figure 2 The architecture described is used for the transmission of secret information during the image animation process.

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

[0050] (3) The size of the secret information is m×32×32. The size of the steganographic image is 256×256. The values ​​of m are 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), the number of secret information bits per pixel can be calculated to be 0.0625, 0.125, 0.25, 0.5 and 1 respectively. An example of the generated steganographic image is shown below. Figure 4As shown, "original image" represents the content image, "animated image" represents the result of only animation without secret data, and "steganographic image" represents the result of animation with embedded secret data; the embedding capacities corresponding to (a)-(e) are 4096, 8192, 16384, 32768 and 65536 bits, respectively.

[0051]

[0052] (4) For comparison, we used four popular steganalysis algorithms: SUNIWARD (spatial universal wavelet relative distortion), WOW (wavelet weighting), HILL (high-pass, low-pass, low-pass), and DFEI (enhanced image steganalysis function), embedding secret data into images that underwent style transfer using only an embedding network. The embedding capacity was set to t = [4096, 8192, 16384, 32768, 65536]. To test the undetectability of our method, we employed four popular steganalysis schemes: SPAM (subtracting pixel adjacency matrix), SRMQ1 (single quantization step spatial enrichment model), maxSRMd2 (channel-aware maximum displacement enrichment model), and SCRMQ1 (quantization step spatial color enrichment model). Half of the carrier image and half of the steganalyte image were used for training, and the remainder for testing. The undetectability evaluation criterion was the minimum steganalysis error P obtained on the test set under the same prior conditions. E , defined in equation (7). P FA and P MD These represent the false alarm rate and the false negative rate, respectively. A higher P... E This implies higher undetectability. The undetectability results of this invention compared to the SUNIWARD, WOW, HILL, and DFEI schemes, such as... Figure 5 As shown, the steganography algorithm of this invention for style-transferred images demonstrates superior undetectability compared to existing steganography algorithms.

[0053]

[0054] If the aforementioned functions are implemented as 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 this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0055] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for steganography during image animation, characterized in that, Includes the following steps: Construct a network architecture for information hiding, the network architecture including a generator, a decoder and a discriminator; Adversarial training is performed based on the loss functions of the generator, decoder, and discriminator to obtain the trained generator, decoder, and discriminator. The sending end inputs a natural world image and the information to be hidden into a trained generator to obtain an animated steganographic image. The generator includes an animation module, an information acquisition module, and a steganographic 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 concatenated in the channel dimension and then input into the steganographic block to output an animated steganographic image. The sending end transmits the animated steganographic image through a public channel; The receiving end acquires the animated steganographic image through a public channel, inputs the animated steganographic image into the trained decoder, and the decoder outputs the information to be hidden. The generator includes inputs on both sides. One side is a natural world image, and the other side is the information to be hidden. The natural world image input side structure includes: two convolutional blocks, a downsampling block, a convolutional block, a depth splitting block, a downsampling block, a convolutional block, eight inverted residual blocks, a convolutional block, an upsampling block, a depth splitting block, a convolutional block, an upsampling block, two convolutional blocks, a standard convolutional block and a Tamh activation block, and a hybrid convolutional block. The information to be hidden input side structure includes a steganalysis upsampling block. The outputs on both sides are connected in the channel dimension and then input to the steganalysis block to output an animated steganalysis image. The convolutional block includes standard convolution, instance normalization, and LRelu activation; the upsampling block includes upsampling and depth separation blocks; the steganography upsampling includes three sets of transposed convolutions plus LRelu activation plus instance normalization; the hybrid convolutional block includes standard convolution, LRelu activation, and instance normalization; the depth separation block includes depth-separable convolution, instance normalization, LRelu activation, and convolutional blocks; the inverted residual block includes convolutional blocks, depth-separable convolution, instance normalization, LRelu activation, standard convolution, and instance normalization; and the downsampling block includes downsampling and two depth separation blocks.

2. The method for steganography during image animation according to claim 1, characterized in that, The generator and decoder are pre-stored at the sending and receiving ends.

3. The method for steganography during image animation according to claim 1, characterized in that, The steganographic block consists of three sets of standard convolutions plus LRelu activation plus instance normalization.

4. The method for steganography during image animation according to claim 1, characterized in that, The decoder takes an animated steganographic image as input and outputs the information to be hidden. The decoder includes three sets of standard convolutions.

5. A method for steganography during image animation according to claim 1, characterized in that, The discriminator takes either an animated steganographic image or a real animated image as input and outputs whether the image is an animated steganographic 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.

6. A method for steganography during image animation according to claim 1, characterized in that, The total loss function of the generator and decoder for: in, For the generator's processing, This refers to the decoding process. This refers to the processing procedure of the discriminator. Real-world images This is a grayscale image of the animated image. , , , , The values ​​are 10, 1.5, 3, 30, and 1 respectively. For content loss, For grayscale style loss, Loss due to color reconstruction.

7. A method for steganography during image animation according to claim 1, characterized in that, The total loss function of the discriminator for: in, For animated images, This is a smoothed grayscale image of the animation.

8. A method for steganography during image animation according to claim 1, characterized in that, The adversarial training based on the loss functions of the generator, decoder, and discriminator includes the following specific training process: 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 real-world images as content images, inputting the content images and secret information into the generator to obtain steganographic images, and simultaneously inputting the steganographic images and real animation images into the discriminator for adversarial training.

Images