Carrier image enhancement method based on generative network

A carrier image, generative technology, applied in the field of information hiding, which can solve the problems of weak mobility, low security, and deceiving non-target steganalyzers.

CN114827379APending Publication Date: 2022-07-29SICHUAN UNIV
4 Cites 1 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Publication Date
2022-07-29

Smart Images

  • Figure 1
    Figure 1
  • Figure 2
    Figure 2
  • Figure 3
    Figure 3
Patent Text Reader

Abstract

A carrier image enhancement method based on a generative network comprises the steps of training an adversarial noise generator: inputting a carrier image in a training set into the adversarial noise generator to obtain adversarial noise; adding the anti-noise into the carrier image to obtain an enhanced carrier image; generating steganography noise by using an adaptive steganography algorithm, and adding the steganography noise into the enhanced carrier image to obtain an enhanced secret-carrying image; the mean square error between the carrier image and the enhanced secret-carrying image is used as mean square error loss; inputting the enhanced secret-carrying image into a pre-trained steganography analyzer to obtain a prediction result; taking the cross entropy loss between the prediction result and the real label of the carrier image as the adversarial loss of an adversarial noise generator; performing weighted summation on the mean square error loss and the adversarial loss to obtain total loss; and through back propagation, an adam optimization method is utilized to optimize the anti-noise generator. Compared with an existing carrier image enhancement algorithm, the carrier image enhancement algorithm is higher in mobility.
Need to check novelty before this filing date? Find Prior Art

Description

technical field

[0001] The invention relates to the technical field of information hiding, in particular to a carrier image enhancement method based on a generative network. Background technique

[0002] As one of the important technologies in the field of information hiding technology, image steganography uses the information redundancy of digital images to hide secret information, which is difficult to be detected. Image steganography hides secret information into normal image carriers, which greatly reduces the risk of malicious interception and tampering, so it is also widely used in the field of intelligence transmission. In recent years, image steganography researchers have proposed many spatially adaptive steganography algorithms that associate the embedded position of secret information with image content, such as HUGO (Highly Undetectable steGO), WOW (Wavelet Obtained Weights), S-UNIWARD (SpatialUniversal) Wavelet Relative Distortion), HILL (High-pass, Low-pass, an...

Examples

Embodiment 1

[0028] A carrier image enhancement method based on a generative network. By adding anti-noise to the carrier image, an enhanced carrier image is obtained. The enhanced carrier image can effectively deceive the steganalyzer and the traditional steganalysis method SRM after embedding secret information.

[0029] A. Training phase (eg figure 1 shown)

[0030] Step 1. Input the carrier image c in the training set into the anti-noise generator G. The anti-noise generator G can choose a fully convolutional network such as FCN and U-Net. In this example, the generator selects the FCN network whose structure is as follows figure 2 shown. The anti-noise generator will obtain the anti-noise n according to the content information of the carrier image.

[0031] Step 2. Add the adversarial noise n to the carrier image c to obtain the enhanced adversarial image adv_c.

[0032] Step 3. Use S-UNIWARD as the adaptive steganography algorithm in this example, generate steganographic noise s...

Embodiment 2

[0049] In order to improve the mobility, the present invention also considers a mobility improvement strategy based on a dual-target steganalyzer, according to the differences in the mobility of the anti-noise generators obtained when training with different target steganalyzers. Divide the steganalyzers into two groups, and then extract a steganalyzer from the two groups to combine, and jointly train as the target steganalyzer to provide adversarial loss, and the rest is the same as when attacking a single steganalyzer. keep it the same. Using two different steganalysis networks as the target steganalyzer can help the adversarial noise generator better learn the differences between the two different steganalyzers and find the common weaknesses between the two models, Generate more mobile adversarial noise.

[0050] A. Training phase (eg image 3 shown)

[0051] Specifically, steps 1 to 6 are the same as those in Embodiment 1.

[0052] Step 7. When implementing the mobilit...