CN114782697A - Adaptive steganography detection method for confrontation sub-field - Google Patents

Adaptive steganography detection method for confrontation sub-field Download PDF

Info

Publication number
CN114782697A
CN114782697A CN202210467768.4A CN202210467768A CN114782697A CN 114782697 A CN114782697 A CN 114782697A CN 202210467768 A CN202210467768 A CN 202210467768A CN 114782697 A CN114782697 A CN 114782697A
Authority
CN
China
Prior art keywords
domain
sub
field
loss
adaptive
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202210467768.4A
Other languages
Chinese (zh)
Other versions
CN114782697B (en
Inventor
王宏霞
章蕾
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Sichuan University
Original Assignee
Sichuan University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Sichuan University filed Critical Sichuan University
Priority to CN202210467768.4A priority Critical patent/CN114782697B/en
Publication of CN114782697A publication Critical patent/CN114782697A/en
Application granted granted Critical
Publication of CN114782697B publication Critical patent/CN114782697B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computational Linguistics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Evolutionary Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a method for detecting adaptive steganography in a confrontation sub-field, which aims at the problem of carrier source mismatch in image steganography detection. The sub-domain adaptive classifier Y minimizes the sub-domain adaptive loss while classifying to reduce the difference of inter-domain related sub-domain feature distribution. The method is simple in calculation and easy to implement, can effectively improve the accuracy of the deep learning-based steganography detection model under the condition of carrier source mismatch, and has practical value.

Description

Adaptive steganography detection method for antagonist field
Technical Field
The invention relates to the technical field of information hiding, in particular to a method for detecting adaptive steganography in the field of confrontation.
Background
The image steganography embeds the secret information into the image, so that the secret information content and the communication behavior can be hidden. The purpose of steganography detection is to detect secret communications established through steganography. In recent years, many high-performance image steganography detection methods based on deep learning emerge. The literature (Mehdi Borouund, Mo Chen, Jessica Fridrich, IEEE Transactions on Information principles and Security,14(5),1181-1193) proposes an end-to-end depth residual network SRNet for steganography detection. When steganographic detection models trained on one carrier source are used to detect images from different carrier sources, the detection error of the steganographic detection models is often increased due to mismatch between the two carrier sources, which is referred to as the carrier source mismatch problem. When steganography detection is carried out, it is very difficult to obtain information such as an original carrier image, a steganography algorithm, an embedding rate and the like adopted by a steganographer for steganography, that is, differences between a training set and a test set are inevitable, which is also a reason why a steganography detection model is difficult to deploy in practical application. The carrier source mismatch problem and the domain adaptation problem in steganography detection are quite similar. The domain adaptive methods can be divided into two categories, one is a metric-based domain adaptive method. The method comprises the steps of firstly mapping the characteristics of a source field and a target field into a high-dimensional characteristic space, then measuring the distance between the mappings in the high-dimensional characteristic space based on certain difference measurement, and realizing the distribution alignment of the source field and the target field by minimizing the measurement index of the inter-domain distribution difference. The literature (Ghifary M, Kleijn W B, Zhang M, Pacific Rim International Conference on Artificial Intelligence,2014: 898-. The MMD adaptation layer added behind the feature layer is trained to reduce the mismatching of the distribution of the inter-domain features, and most of the follow-up research is developed based on the idea. Yet another approach is to learn a representation that can both classify and not distinguish which domain is by selecting an antagonistic loss to minimize inter-domain distance. In the literature (Ganin Y, Ustinova E, Ajakan H, et al, the Journal of machine Learning Research,2016,17(1): 2096-. In this work, the learning objectives of the network are: the extracted features enable the classifier to classify the sample correctly as much as possible, and enable the discriminator not to distinguish whether the sample comes from the source field or the target field. For both of the above-mentioned domain adaptive methods, the domain adaptive method using only counterlearning is insufficient to minimize the distance of inter-domain feature distribution, and the method using only minimization metric lacks the learning of domain-independent features. Aiming at the problem of carrier source mismatch in image steganography detection, the method utilizes test set information, uses a domain self-adaption method to solve the problem of carrier source mismatch, takes training set data as a source domain, takes test set data as a target domain, and improves the detection performance of a steganography detection model in the target domain by minimizing the characteristic distribution difference between the source domain and the target domain.
Disclosure of Invention
In order to overcome the defects in the prior art and improve the accuracy of the steganography detection model under the mismatch condition, the invention provides a method for adapting the sub-domain, which can enable the characteristics generated by the steganography detection model in the source domain and the target domain to be more similar, reduce the difference of the inter-domain characteristic distribution and improve the classification accuracy of the steganography detection model in the target domain.
The technical scheme for realizing the invention is as follows:
a countercheck sub-field self-adaptive steganography detection method adopts a countercheck learning strategy, and performs countercheck learning with a steganography feature extractor F through a field discriminator D, so that features generated by the steganography feature extractor F in a source field and a target field are more similar. The sub-domain adaptive classifier Y minimizes the sub-domain adaptive loss while classifying to reduce the difference of inter-domain related sub-domain feature distribution. Obtaining a steganography detection model MJ for classifying in a target field by simultaneously minimizing classification loss, countermeasure loss and sub-field self-adaptive loss; the method comprises the following main steps:
step 1, taking a steganography detection model which is trained on a source field as a pre-training model M; firstly, removing a classification layer of a pre-training model M to serve as a steganographic feature extractor F; tagged Source Domain data and non-taggedThe target domain data of the label can obtain the characteristics after passing through the l layer of the steganographic characteristic extractor F
Figure BDA0003625150410000021
And
Figure BDA0003625150410000022
step 2, the characteristics obtained in the step 1 are simultaneously sent to a classification branch and a judgment branch; in the classification branch, the network structure of the sub-domain adaptive classifier Y is the same as the classification layer of the pre-training model M, and the parameters are initialized. Detection characteristics f of sub-domain self-adaptive classifier Y according to source domain steganographyslAnd target domain steganography detection feature ftlRespectively giving predicted values, and simultaneously calculating classification loss L1 and sub-domain adaptive loss L2:
Figure BDA0003625150410000023
in the formula nsThe number of samples in the source field; classification loss L1 as classifier predictor
Figure BDA0003625150410000024
And source domain real label
Figure BDA0003625150410000025
Cross entropy loss between.
Figure BDA0003625150410000026
Wherein C is the label category, C belongs to {0,1}, and C is the number of label categories; n issIs the number of samples in the source field, ntThe number of samples in the target field is; w is the weight of the sample belonging to class c;
Figure BDA0003625150410000027
representing the inner product of the features.
In the discrimination branch, the domain discriminator D discriminates the features according to the input
Figure BDA0003625150410000028
A label of the prediction domain is given, i.e. from which domain the feature is determined. n is a radical of an alkyl radicaltCalculating a domain prediction label for the number of target domain samples
Figure BDA0003625150410000029
And a field truth label diAs the countering loss L3:
Figure BDA00036251504100000210
step 3, calculating the total loss function
L＝L1+λL2+ωL3
λ and ω are trade-off parameters for sub-domain adaptive loss and counterloss, respectively; in the training process of the model, the parameters of the steganographic feature extractor F, the sub-domain adaptive classifier Y and the domain discriminator D are updated by minimizing the total loss and performing back propagation, so that the difference of the feature distribution of the source domain and the target domain is reduced.
And 4, stopping training when the set upper limit of the training times is reached, and combining the steganographic feature extractor F and the sub-field self-adaptive classifier Y to obtain a steganographic detection model MJ for classifying in the target field.
Further, the network structure of the domain discriminator D in step 2 is not unique, and can be adjusted according to actual conditions.
Furthermore, the method can be applied to spatial domain steganography detection and JPEG domain steganography detection.
Compared with the prior art, the invention has the advantages that,
1. the method adopts a confrontation learning strategy, and confronts with the steganographic feature extractor by introducing a domain discriminator, and requires that the source domain features and the target domain features generated by the steganographic feature extractor are similar as much as possible.
2. And reducing the distribution difference of the source domain characteristic and the target domain characteristic by calculating and minimizing the adaptive loss of the sub-domain.
Drawings
FIG. 1 is a schematic flow chart of the present invention.
FIG. 2 is a block diagram of the sub-domain adaptive steganography detection method of the present invention.
Fig. 3 is a graph comparing experimental results of the present invention on ucidv.2 data set with a prior method.
Detailed Description
The invention is further described with reference to the following figures and specific examples.
The present invention addresses the problem of carrier source mismatch in image steganography detection. The mismatch of the carrier source comprises various conditions, such as data set mismatch, steganographic algorithm mismatch, embedding rate mismatch, QF mismatch, image processing process mismatch and the like.
The method comprises the following specific steps as shown in figure 1:
S1：
Figure BDA0003625150410000031
represents a tagged source domain, comprising nsThe individual of the samples having the label thereon,
Figure BDA0003625150410000032
is a sample
Figure BDA0003625150410000033
The tag vector of (2). Since steganography detection is a binary task, C ═ 2.
Figure BDA0003625150410000034
Representing the target domain without the label. And taking the steganography detection model which is trained on the source field as a pre-training model M.
As can be seen in conjunction with fig. 2, the classification layer of the pre-trained model M is first removed as a steganographic feature extractor F; after the source field data with the label and the target field data without the label pass through the steganography feature extractor F, the steganography detection feature F of the source field is obtainedslAnd target domain steganographic detection featuresSign ftlThe obtained characteristics are shared by the sub-domain self-adaptive classifier Y and the domain discriminator D, and are simultaneously sent to the classification branch and the discrimination branch.
S2: the use of domain-adaptive methods that only deal with learning is not sufficient to minimize the distance of inter-domain feature distributions, while methods that use only minimization metrics lack the learning of domain-independent features. Therefore, the present invention proposes a counterincident domain adaptive network for mismatch steganography detection to address these limitations. On one hand, from the perspective of reducing inter-domain differences, the feature distribution of related sub-domains is aligned, the inter-class distance is expanded, and the intra-class distance is reduced. On the other hand, one domain discriminator is constructed to carry out antagonistic learning with the steganography detection model, so that the characteristics of the model generated on two domains are more similar. As shown in fig. 2, the anti-sub-field adaptive network provided by the present invention mainly comprises a steganographic feature extractor F, a sub-field adaptive classifier Y and a field discriminator D, and mainly aims to improve the detection accuracy of the pre-trained steganographic detection model under the mismatch condition. In the classification branch, the network structure of the sub-domain adaptive classifier Y is the same as the classification layer of the pre-training model M, and the parameters are initialized. The sample can obtain the characteristics after passing through the first layer of the steganographic characteristic extractor F
Figure BDA0003625150410000041
And
Figure BDA0003625150410000042
the sub-domain self-adaptive classifier Y gives a predicted value of the source domain image
Figure BDA0003625150410000043
By calculation of
Figure BDA0003625150410000044
And source field real label YsCross entropy loss between them yields a classification loss L1:
Figure BDA0003625150410000045
the classifier Y gives a predicted value of the target field image
Figure BDA0003625150410000046
And source domain feature fslTarget area characteristics ftlTrue source field label YsTogether used to calculate the sub-domain adaptive loss L2:
Figure BDA0003625150410000047
wherein c is a label category, and c belongs to {0,1 }; n issIs the number of source domain samples, ntThe number of samples in the target field is; w is the weight of the sample belonging to class c;
Figure BDA0003625150410000048
representing an inner product of the features; in the formula, a sample xiThe weights belonging to class c are calculated as:
Figure BDA0003625150410000049
yicis a label vector yiThe c-th value of (a).
In the decision branch, the feature first passes through a gradient inversion layer (GRL) to simplify the countermeasure training process. The principle of gradient inversion layer implementation is: the input is kept unchanged in the forward propagation process, and the gradient is multiplied by a negative fixed value mu in the backward propagation process, so that the automatic inversion of the gradient is realized to achieve the aim of resisting training. The effect of the gradient inversion layer in forward and backward propagation can be expressed using two pseudo-functions:
forward propagation:
Rμ(x)＝x
and (3) back propagation:
Figure BDA00036251504100000410
in the formula, I is an identity matrix. The domain discriminator D gives a label of the prediction domain based on the inputted feature, i.e., determines from which domain the feature is coming. n istCalculating a domain prediction label for the number of target domain samples
Figure BDA0003625150410000051
And a field truth label diAs the countering loss L3:
Figure BDA0003625150410000052
s3: calculating a total loss function
L＝L1+λL2+ωL3
λ and ω are trade-off parameters for sub-domain adaptive loss and counterloss, respectively; in the training process of the model, parameters of a steganographic feature extractor F, a sub-domain adaptive classifier Y and a domain discriminator D are updated by minimizing total loss and performing back propagation, so that the difference of feature distribution of a source domain and a target domain is reduced;
s4: and when the set upper limit of the training times is reached, stopping training, and combining the steganographic feature extractor F and the sub-field self-adaptive classifier Y to obtain a steganographic detection model MJ for classifying in the target field.
The effectiveness of the process is illustrated below with reference to a specific experiment
Given a pre-training model SRNet trained on a BOSSBase data set by using an S-UNIWARD steganography algorithm with an embedding rate of 0.4bpp (bits per pixel, bpp), when detecting S-UNIWARD steganography image pairs on other data sets, such as a UCIDv.2 data set, due to the difference between the training set and the testing set, the detection accuracy of the pre-training model may be reduced to different degrees, which belongs to the data set mismatch in the problem of carrier source mismatch. In order to verify the effectiveness of the invention, data set mismatch experiments are carried out on three steganographic algorithms of S-UNIWARD, WOW and HUGO. And randomly selecting 500 images on the BOSSBase data set as carrier images, and performing steganography to obtain 500 secret-carrying images. The 500 pairs of images and labels are taken as source fields. 500 images are randomly selected from a UCIDv.2 data set to serve as carrier images, and 500 carrier images are obtained after steganography is carried out. The 500 pairs of images are taken as the target area.
Step 1, firstly, removing a classification layer of an SRNet pre-training model to serve as a steganography feature extractor F; after the source field data with the label and the target field data without the label pass through the steganography feature extractor F, the steganography detection feature F of the source field is obtainedslAnd target domain steganography detection feature fslAnd simultaneously sending the obtained characteristics to the classification branch and the judgment branch.
And 2, enabling the network structure of the sub-field self-adaptive classifier Y to be the same as the classification layer of the SRNet pre-training model, wherein the network structure is a full connection layer, and the parameters of the full connection layer are initialized. The sub-domain self-adaptive classifier Y gives a predicted value of the source domain image
Figure BDA0003625150410000053
By calculation of
Figure BDA0003625150410000054
And source field real label YsThe cross entropy loss between the two results in a classification loss L1, and the classifier Y gives a predicted value of the target field image
Figure BDA0003625150410000055
And source domain feature fslTarget area characteristics ftlTrue source field label YsTogether used to calculate the sub-domain adaptation loss L2. The domain discriminator D consists of three linear layers to predict the domain label of the feature. The feature dimension of SRNet is 512, and the hidden layer dimension of the domain discriminator D is set to 512. The domain discriminator D gives a label of the prediction domain based on the input features, and predicts the label by calculating the domain
Figure BDA0003625150410000061
And a field truth label diAs the counter-measure loss L3.
And 3, calculating and minimizing a total loss function, performing back propagation, and updating parameters of the steganographic feature extractor F, the sub-domain adaptive classifier Y and the domain discriminator D.
And 4, stopping training after 200 epochs are trained. And combining the steganographic feature extractor F and the sub-field self-adaptive classifier Y to obtain a new SRNet model for classifying in the target field.
As can be seen from the data set mismatch experimental result in FIG. 3, by performing field self-adaptation on the SRNet model by using the method of the present invention, the detection accuracy of the model under the mismatch condition can be improved by 9.9% -10.9%, and the method is effective when three different steganographic algorithms are used, and has practical value.

Claims (2)

1. A countercheck sub-field self-adaptive steganography detection method adopts a countercheck learning strategy, and performs countercheck learning with a steganography feature extractor F through a field discriminator D, so that the features generated by the steganography feature extractor F in a source field and a target field are more similar; the sub-domain adaptive classifier Y minimizes the sub-domain adaptive loss to reduce the difference of the inter-domain related sub-domain feature distribution while classifying; obtaining a steganography detection model MJ for classifying in a target field by simultaneously minimizing classification loss, countermeasure loss and sub-field self-adaptive loss; the method comprises the following main steps:
step 1, taking a steganography detection model which is trained in a source field as a pre-training model M; firstly, removing a classification layer of a pre-training model M to serve as a steganographic feature extractor F; the source domain data with labels and the target domain data without labels can obtain the characteristics after passing through the l layer of the steganographic characteristic extractor F
Figure FDA0003625150400000011
And
Figure FDA0003625150400000012
step 2, the characteristics obtained in the step 1 are simultaneously sent to a classification branch and a judgment branch; in the classification branchThe network structure of the sub-domain adaptive classifier Y is the same as the classification layer of the pre-training model M, the parameters are initialized, and the sub-domain adaptive classifier Y detects the characteristics f according to the steganography of the source domainslAnd target domain steganography detection feature ftlRespectively giving predicted values, and simultaneously calculating classification loss L1 and sub-domain adaptive loss L2:
Figure FDA0003625150400000013
in the formula nsThe number of samples in the source field; classification loss L1 is classifier predictor
Figure FDA0003625150400000014
And source field real label
Figure FDA0003625150400000015
Cross entropy loss between;
Figure FDA0003625150400000016
wherein C is the label category, C belongs to {0,1}, and C is the number of label categories; n issIs the number of samples in the source field, ntThe number of samples in the target field is obtained; w is the weight of the sample belonging to class c;
Figure FDA0003625150400000017
representing an inner product of the features;
in the discrimination branch, the domain discriminator D discriminates the features according to the input
Figure FDA0003625150400000018
Giving a label of a prediction domain, namely judging which domain the feature comes from; n istCalculating a domain prediction label for the number of target domain samples
Figure FDA0003625150400000019
And a field truth label diAs the countering loss L3:
Figure FDA00036251504000000110
step 3, calculating the total loss function
L＝L1+λL2+ωL3
Lambda and omega are balance parameters of the sub-field self-adaptive loss and the antagonistic loss respectively; in the training process of the model, parameters of a steganographic feature extractor F, a sub-domain adaptive classifier Y and a domain discriminator D are updated by minimizing total loss and performing back propagation, so that the difference of feature distribution of a source domain and a target domain is reduced;
and 4, stopping training when the set upper limit of the training times is reached, and combining the steganographic feature extractor F and the sub-field self-adaptive classifier Y to obtain a steganographic detection model MJ for classifying in the target field.
2. The method as claimed in claim 1, wherein the network structure of the domain discriminator D in step 2 is not unique and can be adjusted according to actual conditions.
CN202210467768.4A 2022-04-29 2022-04-29 Self-adaptive steganography detection method for anti-domain Active CN114782697B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210467768.4A CN114782697B (en) 2022-04-29 2022-04-29 Self-adaptive steganography detection method for anti-domain

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210467768.4A CN114782697B (en) 2022-04-29 2022-04-29 Self-adaptive steganography detection method for anti-domain

Publications (2)

Publication Number Publication Date
CN114782697A true CN114782697A (en) 2022-07-22
CN114782697B CN114782697B (en) 2023-05-23

Family

ID=82435691

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210467768.4A Active CN114782697B (en) 2022-04-29 2022-04-29 Self-adaptive steganography detection method for anti-domain

Country Status (1)

Country Link
CN (1) CN114782697B (en)

Citations (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130208941A1 (en) * 2012-02-01 2013-08-15 Qingzhong Liu Steganalysis with neighboring joint density
CN106920206A (en) * 2017-03-16 2017-07-04 广州大学 A kind of steganalysis method based on confrontation neutral net
US20190287515A1 (en) * 2018-03-16 2019-09-19 Microsoft Technology Licensing, Llc Adversarial Teacher-Student Learning for Unsupervised Domain Adaptation
CN111598762A (en) * 2020-04-21 2020-08-28 中山大学 Generating type robust image steganography method
CN111696021A (en) * 2020-06-10 2020-09-22 中国人民武装警察部队工程大学 Image self-adaptive steganalysis system and method based on significance detection
CN112132738A (en) * 2020-10-12 2020-12-25 中国人民武装警察部队工程大学 Image robust steganography method with reference generation
CN112634117A (en) * 2020-12-22 2021-04-09 中山大学 End-to-end JPEG domain image steganography method based on generation countermeasure network
CN112907431A (en) * 2021-02-26 2021-06-04 中国科学技术大学 Steganalysis method for resisting steganography robustness
CN113538202A (en) * 2021-08-05 2021-10-22 齐鲁工业大学 Image steganography method and system based on generative steganography confrontation
CN114065861A (en) * 2021-11-17 2022-02-18 北京工业大学 Domain self-adaption method and device based on contrast counterwork learning
CN114169462A (en) * 2021-12-14 2022-03-11 四川大学 Feature-guided depth sub-field self-adaptive steganography detection method
CN114257697A (en) * 2021-12-21 2022-03-29 四川大学 High-capacity universal image information hiding method

Patent Citations (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130208941A1 (en) * 2012-02-01 2013-08-15 Qingzhong Liu Steganalysis with neighboring joint density
CN106920206A (en) * 2017-03-16 2017-07-04 广州大学 A kind of steganalysis method based on confrontation neutral net
US20190287515A1 (en) * 2018-03-16 2019-09-19 Microsoft Technology Licensing, Llc Adversarial Teacher-Student Learning for Unsupervised Domain Adaptation
CN111598762A (en) * 2020-04-21 2020-08-28 中山大学 Generating type robust image steganography method
CN111696021A (en) * 2020-06-10 2020-09-22 中国人民武装警察部队工程大学 Image self-adaptive steganalysis system and method based on significance detection
CN112132738A (en) * 2020-10-12 2020-12-25 中国人民武装警察部队工程大学 Image robust steganography method with reference generation
CN112634117A (en) * 2020-12-22 2021-04-09 中山大学 End-to-end JPEG domain image steganography method based on generation countermeasure network
CN112907431A (en) * 2021-02-26 2021-06-04 中国科学技术大学 Steganalysis method for resisting steganography robustness
CN113538202A (en) * 2021-08-05 2021-10-22 齐鲁工业大学 Image steganography method and system based on generative steganography confrontation
CN114065861A (en) * 2021-11-17 2022-02-18 北京工业大学 Domain self-adaption method and device based on contrast counterwork learning
CN114169462A (en) * 2021-12-14 2022-03-11 四川大学 Feature-guided depth sub-field self-adaptive steganography detection method
CN114257697A (en) * 2021-12-21 2022-03-29 四川大学 High-capacity universal image information hiding method

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
DENIS VOLKHONSKIY等: "Steganographic generative adversarial networks", TWELFTH INTERNATIONAL CONFERENCE ON MACHINE VISION (ICMV 2019) *
LEI ZHANG等: "Dataset mismatched steganalysis using subdomain adaptation with guiding feature", TELECOMMUNICATION SYSTEMS *
才智等: "生成式对抗网络应用领域综述", 上海电力大学学报 *
陈君夫等: "基于深度学习的图像隐写分析综述", 软件学报 *

Also Published As

Publication number Publication date
CN114782697B (en) 2023-05-23

Similar Documents

Publication Publication Date Title
CN110837850B (en) Unsupervised domain adaptation method based on counterstudy loss function
CN107563155B (en) Security steganography method and device based on generation of countermeasure network
Wu et al. Genetic algorithm based methodology for breaking the steganalytic systems
CN111242157A (en) Unsupervised domain self-adaption method combining deep attention feature and conditional opposition
CN113076994B (en) Open-set domain self-adaptive image classification method and system
CN107240061B (en) Watermark embedding and extracting method and device based on dynamic BP neural network
CN113221903B (en) Cross-domain self-adaptive semantic segmentation method and system
CN112966740B (en) Small sample hyperspectral image classification method based on core sample self-adaptive expansion
WO2024051183A1 (en) Backdoor detection method based on decision shortcut search
CN114257697B (en) High-capacity universal image information hiding method
CN111726472A (en) Image anti-interference method based on encryption algorithm
Meng et al. High-capacity steganography using object addition-based cover enhancement for secure communication in networks
CN110889797A (en) Depth self-adaptive image hiding method based on confrontation sample generation
CN112907431B (en) Steganalysis method for robust countersteganalysis
CN113362216A (en) Deep learning model encryption method and device based on backdoor watermark
Kaur et al. Feature selection using mutual information and adaptive particle swarm optimization for image steganalysis
Sultan et al. A new framework for analyzing color models with generative adversarial networks for improved steganography
Megías et al. Subsequent embedding in targeted image steganalysis: Theoretical framework and practical applications
CN114782697A (en) Adaptive steganography detection method for confrontation sub-field
CN113011513A (en) Image big data classification method based on general domain self-adaption
Goodman Transferability of adversarial examples to attack cloud-based image classifier service
CN114169462A (en) Feature-guided depth sub-field self-adaptive steganography detection method
Wang et al. When not to classify: Detection of reverse engineering attacks on DNN image classifiers
CN113159317A (en) Antagonistic sample generation method based on dynamic residual corrosion
Zhang et al. Technology of image steganography and steganalysis based on adversarial training

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant