CN108805161B - Steganography detection method for multi-embedding-rate encrypted image - Google Patents
Steganography detection method for multi-embedding-rate encrypted image Download PDFInfo
- Publication number
- CN108805161B CN108805161B CN201810375675.2A CN201810375675A CN108805161B CN 108805161 B CN108805161 B CN 108805161B CN 201810375675 A CN201810375675 A CN 201810375675A CN 108805161 B CN108805161 B CN 108805161B
- Authority
- CN
- China
- Prior art keywords
- training
- images
- efld
- image
- classifier
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
Abstract
The invention relates to a steganography detection method of a multi-embedding-rate encrypted image. The method obtains two integrated FLD classifiers through image training with different embedding rates. The first integrated FLD classifier is obtained by training low embedding rate images and carrier images, training samples which better represent the characteristics of the carrier images and the steganographic images can be selected, and the selected samples and the high embedding rate versions of the selected samples are trained together to obtain the second integrated FLD classifier. The main steps in classifying the unknown class of images are: the first integrated FLD classifier classifies images with the number of votes in a certain range, the second integrated FLD classifier classifies the rest images with the number of votes in the certain range, and finally the images which are not classified are classified by the first integrated FLD classifier. The method of the invention respectively utilizes the advantages of the two integrated FLD classifiers, thereby improving the identification accuracy of the JPEG steganographic image.
Description
Technical Field
The invention relates to a steganography detection method of a multi-embedding-rate dense image, aiming at the characteristic that a plurality of embedding-rate steganography images can be obtained during training, and the identification accuracy rate of a low-embedding-rate image is improved through a high-embedding-rate image.
Background
Steganalysis is an important component in multimedia information security, the main task of which is to detect whether secret information is present in a digital carrier. Digital images are most commonly used in a large number of digital carriers because a 512 x 512 size digital image can carry enough secret information. Digital images can be classified into many categories due to their formats, and JPEG (Joint Picture Expert Group Joint photographic experts Group) images are widely used in life because of their compression and image reconstruction effects. Many steganography methods are designed on the basis of JPEG images, and the steganography images containing secret information are mixed in a large number of normal JPEG images of the Internet for information transmission, so that the steganography method has high concealment. The current excellent JPEG image steganography method is characterized in that: the statistical properties are unchanged and the amount of information embedding is sufficient. If the embedding rate of the embedded information is 0.05bpac (bit per non-zero AC DCT coeffient bit per non-zero discrete cosine transform AC coefficient), 0.1bpac and other low embedding rates, the steganographic analysis has low recognition accuracy rate on the excellent JPEG image steganographic method. One research content of steganalysis is that steganalysis characteristics capable of better distinguishing steganalysis images from normal digital images are extracted by grasping the difference between the steganalysis images and the principle of steganalysis algorithm, and the other research content is that a classifier which is in accordance with application scenes is designed.
Disclosure of Invention
The invention aims to provide a steganography detection method of a multi-embedding-rate dense image, which aims at the defects of the prior art, can obtain steganography image feature sets with different embedding rates when a classifier is trained by steganography analysis, designs a new integrated FLD (Fisher linear discriminant) classifier, and improves the identification accuracy of a low-embedding-rate image by using a high-embedding-rate steganography image.
In order to achieve the purpose, the invention adopts the following technical scheme:
1. a steganography detection method of a multi-embedding-rate encrypted image is characterized by comprising the following operation steps:
(1) is low inEFLD is obtained by training embedding rate image and carrier image 1 (integrated fisher linear discriminant classifier 1): training carrier images and low embedding rate images to obtain an integrated FLD classifier EFLD 1 And through EFLD 1 Selecting training images with high ticket number;
(2) training low embedding rate image, high embedding rate image and carrier image to obtain EFLD 2 (integrated fisher linear discriminant classifier 2): obtaining a second integrated FLD classifier EFLD through the training of the training image selected in the last step and the high embedding rate version thereof 2 ；
(3) Cross-validation determines parameter values: determining the size of a parameter used for finally classifying the unclassified image through cross validation, namely determining the range of the number of votes;
(4) classifying the unclassified images: first using EFLD 1 Classifying the test images within the range of the number of votes obtained, and using EFLD to the rest test images 2 Classifying the images within the range of the number of votes obtained, wherein the test images which are not classified are obtained by EFLD 1 Classification
Training the low embedding rate image and the carrier image to obtain the EFLD in the step (1) 1 ：
First, there are 3 nxp steganalysis data sets, where N is the number of samples, p is the dimension of the samples, and N is the number of integrated classifier sub-classifiers. The first data set is a carrier image steganalysis feature set, the second data set is a low embedding rate dense image feature set, and the third data set is a high embedding rate dense image feature set. Generating a group of random positive integer sets with the length equal to 0.5n and the maximum value equal to n, wherein the numbers are not repeated, the numbers are used as serial numbers to extract training samples from the three data sets, and the rest data are used as test samples; after the processing, 0.5n carrier images, 0.5n low-embedding-rate images and 0.5n high-embedding-rate images are in training; a part of training images are required to be left as a cross validation set before a first integrated FLD classifier is trained, and the rest training sets are trained to obtain the first integrated FLD classifier;
when training the first integrated FLD classifier, sample features are extracted randomlySubspace, namely, randomly selecting a certain dimensionality from the original p-dimensional characteristics of the sample; after random subspaces of the carrier image and the low embedding rate image samples are extracted, the optimal projection direction of the FLD (flash light emitting diode) sub-classifier can be calculated
(ii) a The optimal projection direction can be obtained by the optimized calculation according to the formula (1)
The following alternatives may also be used: [1]Calculating an intra-class divergence matrix: respectively calculating the mean value of each characteristic value of the two types of samples, and obtaining the mean value difference of the two types of characteristics; and subtracting the mean value of the characteristic value of each type of characteristic, multiplying the mean value by the transposed matrix of the calculation result, and adding the calculation results of the two types of characteristics to obtain an intra-type divergence matrix. [2]Two types of mean difference matrices are calculated: the matrix is actually [1 ]]And (4) calculating to obtain two types of mean difference matrixes. [3]Adding an L2 optimization term: adding one to the intra-class divergence matrix of the same size as its matrix and multiplying by 10 -10 The identity matrix of (2). [4]Calculating to obtain the optimal projection direction
: the similar inner divergence matrix of the L2 optimization term is added to be divided by the two types of mean difference matrixes to obtain the approximate optimal projection direction
。
Wherein
Is a matrix of divergence within a class,
is a matrix of inter-class divergence,
is shown as
The mean value of the class or classes,
is shown as
Class I the first
The feature vector of each image is calculated,
represents the mean of all training data. Obtaining the best projection direction
Then, calculating a judgment threshold b; the specific calculation method is that the training sample is multiplied by
And obtaining a projection result, taking b as the minimum value of the projection result, and then continuously increasing the size of b, wherein when the classification accuracy is the maximum, the b is the judgment threshold.
Repeating for N times to obtain an integrated FLD classifier EFLD with N sub-classifiers in total 1 (ii) a Using EFLD 1 Classifying the training samples to obtain voting results of all training images; training images with a number of votes within a certain range are extracted from the training imagesHigh embedding rate images corresponding to the training images; the extracted images have higher ticket numbers, which indicates that the judgment of more sub-classifiers is the same, so the images can better embody the characteristics of the steganographic images or the carrier images.
And (2) training the low embedding rate image, the high embedding rate image and the carrier image to obtain the EFLD 2 ：
With the high-ticket-count carrier image and low-embedding steganographic image and high-embedding-rate version thereof, a new integrated FLD classifier EFLD is trained 2 But the training process is slightly different from that in the first step; since the number of selected images is generally smaller than the number of training images in the first step, a bootstrap method is first used to resample these small samples to generate a new training set each time the sub-classifier is trained; generally, after each bootstrap, a certain number of samples are not taken, and these samples are used for OOB (out of bag error rate) tests to determine the value of the parameter; during training, a certain number of images with high embedding rate are slowly and continuously added or removed, and a sub-classifier with the best OOB effect is selected in the process; the same circulation is carried out for N times to obtain an integrated FLD classifier EFLD with N sub-classifiers in total 2 。
The step (3) cross-validation determines parameter values:
training to obtain two integrated FLD classifiers EFLD 1 And EFLD 2 Then, an additional parameter used in the test is required to be calculated, and the function of the parameter is to determine the discrimination threshold of the integrated classifier in the test; the specific method is to perform cross validation by using the training samples which do not participate in training the classifier in the step (1) and to obtain the parameter value which enables the cross validation result to be the best.
The step (4) classifies the unclassified image:
first using EFLD 1 Determining the test sample, EFLD 1 Only test samples with a certain range of votes are classified; EFLD using samples with high embedding rate for training 2 The rest of the test samples were subjected toClassifying, namely classifying only the test samples with the votes within a certain range, and directly skipping the classified test samples; after these two steps, there will be some test samples that have not been classified by EFLD 1 Classifying; this is because the optimal decision plane and decision threshold for low embedding rate images and carrier images are in the EFLD 1 So that the remaining unclassified test samples are in the EFLD 1 The highest identification accuracy is achieved;
the final sample prediction consists of three parts, EFLD 1 Discrimination of high-ticket-count samples, EFLD 2 Discrimination and EFLD for high ticket number samples 1 For all the remaining samples, if there is a sample with 0 voting result in the third part, the EFLD is performed 1 Samples are classified randomly, and the number of samples which need to be judged randomly is very small; after the sample is classified, no matter the sample is wrong, the subsequent classifier can not classify the sample, so that the repeated classification of the same sample can not occur in the three parts;
therefore, the newly integrated FLD classifier designed according to the characteristic that the steganographic image has a plurality of embedding rates can improve the accuracy rate of steganographic analysis.
Compared with the prior art, the invention has the following obvious and prominent substantive characteristics and remarkable technical progress: the existing steganalysis technology needs to improve the recognition accuracy of low-embedding-rate images and can acquire high-embedding-rate images when steganalysis is not considered, the method takes the high-embedding-rate images into account, and the high-embedding-rate images are used for improving the recognition accuracy of the low-embedding-rate images.
Drawings
FIG. 1 is a block diagram of the operational procedure of the method of the present invention.
FIG. 2 is a flow chart of the method of the present invention for training a first integrated FLD classifier.
FIG. 3 is a flow chart of a sub-classifier training in a second integrated FLD classifier of the method of the present invention.
FIG. 4 is a classification flow diagram of the method of the present invention.
Detailed Description
The following detailed description of specific embodiments of the invention refers to the accompanying drawings.
The first embodiment is as follows:
referring to fig. 1, the steganography detection method of the multi-embedding-rate dense image is characterized by comprising the following operation steps:
(1) EFLD obtained by training low embedding rate image and carrier image 1 : training carrier images and low embedding rate images to obtain an integrated FLD classifier EFLD 1 And through EFLD 1 Selecting training images with high ticket number;
(2) training low embedding rate image, high embedding rate image and carrier image to obtain EFLD 2 : obtaining a second integrated FLD classifier EFLD through the training of the training image selected in the last step and the high embedding rate version thereof 2 ；
(3) Cross-validation determines parameter values: determining the size of a parameter used for finally classifying the unclassified image through cross validation, namely determining the range of the number of votes;
(4) classifying the unclassified images: first using EFLD 1 Classifying the test images within the range of the number of votes obtained, and using EFLD to the rest test images 2 Classifying the images within the range of the number of votes obtained, wherein the test images which are not classified are obtained by EFLD 1 And (6) classifying.
Example two: the present embodiment is substantially the same as the first embodiment, and is characterized in that:
referring to fig. 2, the step (1) of training the low embedding rate image and the carrier image to obtain the EFLD 1 ：
An integrated FLD classifier with 50 sub-classifiers in total, which are FLD classifiers with L2 optimization, was first trained using 90% training samples.
Training EFLD 1 The specific steps of (a) are shown in fig. 2. There are 3 sets of 10000 × 8000 steganalysis data, 10000 being the number of samples, 8000 being the dimensions of the samples, and 50 being the number of integrated classifier sub-classifiers. The first data set is a carrier image steganography analysis feature set, the second data set is a low embedding rate secret image feature set, and the third data set is a carrier image steganography analysis feature setThe individual data sets are high embedding rate dense image feature sets. Generating a group of random positive integer sets with the length equal to 5000 and the maximum value equal to 10000, wherein the numbers are not repeated, extracting training samples from the three data sets by taking the numbers as serial numbers, and leaving the rest to be used as test samples, wherein after the processing, 5000 carrier images, 5000 low-embedding rate images and 5000 high-embedding rate images participate in training; before training the first integrated FLD classifier, 10% of training sets are required to be left as cross validation sets, and the remaining 90% of training sets are trained to obtain the first integrated FLD classifier;
when a first integrated FLD classifier is trained, firstly, a random 2000-dimensional subspace is extracted from sample features, namely 2000-dimensional features are randomly selected from original 8000-dimensional features of samples; after random subspaces of the carrier image and the low embedding rate image sample are extracted, the optimal projection direction of the FLD sub-classifier can be calculated; repeating the steps for 50 times to obtain an integrated FLD classifier EFLD with a total of 50 sub-classifiers 1 (ii) a Using EFLD 1 Classifying the training samples to obtain voting results of all training images; training images with the number of votes larger than 10 or smaller than-10 and high embedding rate images corresponding to the training images are taken out.
Example three: the present embodiment is substantially the same as the first embodiment, and is characterized in that:
referring to fig. 3, in the step (2), the low embedding rate image, the high embedding rate image and the carrier image are trained to obtain the EFLD 2 ：
After carrier images with high ticket number, low-embedding steganographic images and versions with high embedding rate are obtained, a new integrated FLD classifier is trained, but the training process is slightly different from the previous one; firstly, resampling a small number of samples by adopting a bootstrap method to generate a new training set, and adding a high embedding rate version of each sub-classifier to train on the basis of a high-order training image to obtain a sub-classifier with a better classification effect when each sub-classifier is trained; the specific steps are shown in FIG. 3, selecting a random subspace for the selected training samples and the high embedding rate versions thereof, using 90% of the selected training samples and 100 additional random high embedding rate imagesTraining an FLD (flash memory) sub-classifier, and recording the OOB (object-oriented) effect of the sub-classifier on the remaining 10% of samples; then adding random 100 high embedding rate images and carrying out FLD sub-classifier training, comparing the FLD sub-classifier with the best OOB effect at present, if the OOB effect is better than the current optimal sub-classifier, continuously adding the high embedding rate images, and if not, removing the currently added 100 high embedding rate images and reselecting 100 high embedding rate images for next training; stopping training when the preset total training times are reached, and taking out the sub-classifier with the best OOB effect; the same cycle is carried out for 50 times to obtain an integrated FLD classifier EFLD with a total of 50 sub-classifiers 2 。
Example four: the present embodiment is substantially the same as the first embodiment, and is characterized in that:
the step (3) cross-validation determines parameter values:
training to obtain two integrated FLD classifiers EFLD 1 And EFLD 2 Then, an additional parameter used in the test is required to be calculated, and the function of the parameter is to determine the discrimination threshold of the integrated classifier in the test; the specific method is that 10% of training samples which do not participate in training the classifier in the step (1) are used for cross validation, and the range of the number of votes which enable the cross validation result to be the best is obtained and is in the range of 0.5 and 1]In the meantime.
Example five: the present embodiment is substantially the same as the first embodiment, and is characterized in that:
referring to fig. 4, the step (4) classifies the unclassified image:
the classification step for the unknown image is shown in fig. 4. First using EFLD 1 Determining the test sample, EFLD 1 Only test samples with a certain range of votes are classified; EFLD using samples with high embedding rate for training 2 Classifying the rest test samples, classifying only the test samples with the vote number within a certain range, and directly skipping the classified test samples; after these two steps, there will be some test samples that have not been classified by EFLD 1 Classifying;
the final sample prediction result consists of three partsIs composed of an EFLD 1 Discrimination of high-ticket-count samples, EFLD 2 Discrimination and EFLD for high ticket number samples 1 For all the remaining samples, if there is a sample with 0 voting result in the third part, the EFLD is performed 1 Samples are classified randomly, and the number of samples which need to be judged randomly is very small; after the sample is classified, no matter the sample is wrong, the subsequent classifier can not classify the sample, so that the repeated classification of the same sample can not occur in the three parts;
therefore, the newly integrated FLD classifier designed according to the characteristic that the steganographic image has a plurality of embedding rates can improve the accuracy rate of steganographic analysis.
Claims (4)
1. A steganography detection method of a multi-embedding-rate encrypted image is characterized by comprising the following operation steps:
(1) EFLD obtained by training low embedding rate image and carrier image 1 ：
Training carrier images and low embedding rate images to obtain an integrated FLD classifier EFLD 1 And through EFLD 1 Selecting training images with high ticket number;
(2) training low embedding rate image, high embedding rate image and carrier image to obtain EFLD 2 ：
Obtaining a second integrated FLD classifier EFLD through the training of the training image selected in the last step and the high embedding rate version thereof 2 ；
(3) Cross-validation determines parameter values:
determining the size of a parameter used for finally classifying the unclassified image through cross validation, namely determining the range of the number of votes;
(4) classifying the unclassified images:
first using EFLD 1 Classifying the test images within the range of the number of votes obtained, and using EFLD to the rest test images 2 Classifying the images within the range of the number of votes obtained, wherein the test images which are not classified are obtained by EFLD 1 Classifying;
classifying the unclassified image in step (4) includes the steps of:
first using EFLD 1 Determining the test sample, EFLD 1 Only test samples with a certain range of votes are classified; EFLD using samples with high embedding rate for training 2 Classifying the rest test samples, classifying only the test samples with the vote number within a certain range, and directly skipping the classified test samples; after these two steps, there will be some test samples that have not been classified by EFLD 1 Classifying;
the final sample prediction consists of three parts, EFLD 1 Discrimination of high-ticket-count samples, EFLD 2 Discrimination and EFLD for high ticket number samples 1 For all the remaining samples, if there is a sample with 0 voting result in the third part, the EFLD is performed 1 Samples are classified randomly, and the number of samples which need to be judged randomly is very small; after the sample is classified, no matter the sample is wrong, the subsequent classifier can not classify the sample, so that the repeated classification of the same sample can not occur in the three parts;
therefore, the newly integrated FLD classifier designed according to the characteristic that the steganographic image has a plurality of embedding rates can improve the accuracy rate of steganographic analysis.
2. The steganography detection method of a multi-embedding-rate dense image as claimed in claim 1, wherein: training low embedding rate image and carrier image to obtain EFLD (extended version display device) 1 ：
There are 3 nxp steganalysis data sets, N is the number of samples, p is the dimension of the samples, and N is the number of integrated classifier sub-classifiers; the first data set is a carrier image steganalysis feature set, the second data set is a low embedding rate dense image feature set, and the third data set is a high embedding rate dense image feature set; generating a group of random positive integer sets with the length equal to 0.5n and the maximum value equal to n, wherein the numbers are not repeated, the numbers are used as serial numbers to extract training samples from the three data sets, and the rest data are used as test samples; after the processing, 0.5n carrier images, 0.5n low-embedding-rate images and 0.5n high-embedding-rate images are in training; a part of training images are required to be left as a cross validation set before a first integrated FLD classifier is trained, and the rest training sets are trained to obtain the first integrated FLD classifier;
when a first integrated FLD classifier is trained, a random subspace is firstly extracted from sample features, namely a certain dimensionality is randomly selected from original p-dimensional features of samples; after random subspaces of the carrier image and the low embedding rate image sample are extracted, the optimal projection direction w of the FLD sub-classifier can be calculated; repeating for N times to obtain an integrated FLD classifier EFLD with N sub-classifiers in total 1 (ii) a Using EFLD 1 Classifying the training samples to obtain voting results of all training images; training images with the number of votes within a certain range and high embedding rate images corresponding to the training images are extracted from the image.
3. The steganography detection method of a multi-embedding-rate dense image as claimed in claim 1, wherein: step (2) low embedding rate image, high embedding rate image and carrier image are trained to obtain EFLD 2 ：
After carrier images with high ticket number, low-embedding steganographic images and versions with high embedding rate are obtained, a new integrated FLD classifier is trained, but the training process is slightly different from the previous one; firstly, resampling a small number of samples by adopting a bootstrap method to generate a new training set, and adding a high embedding rate version of each sub-classifier to train on the basis of a high-order training image to obtain a sub-classifier with a better classification effect when each sub-classifier is trained; during training, a certain number of images with high embedding rate are slowly and continuously added or removed, and a sub-classifier with the best OOB effect is selected in the process; after the same circulation is carried out for N times, an integrated FLD classifier EFLD with N sub-classifiers in total is obtained 2 。
4. The steganography detection method of a multi-embedding-rate dense image as claimed in claim 1, wherein: step (3) cross validation determination parameter values:
training to obtain two integrated FLDsClassifier EFLD 1 And EFLD 2 Then, an additional parameter used in the test is required to be calculated, and the function of the parameter is to determine the discrimination threshold of the integrated classifier in the test; the specific method is to utilize the training samples which do not participate in the training of the classifier in the step (1) to carry out cross validation test, and to obtain the parameter value which enables the cross validation result to be the best.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810375675.2A CN108805161B (en) | 2018-04-25 | 2018-04-25 | Steganography detection method for multi-embedding-rate encrypted image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810375675.2A CN108805161B (en) | 2018-04-25 | 2018-04-25 | Steganography detection method for multi-embedding-rate encrypted image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108805161A CN108805161A (en) | 2018-11-13 |
| CN108805161B true CN108805161B (en) | 2022-08-05 |
Family
ID=64093838
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810375675.2A Active CN108805161B (en) | 2018-04-25 | 2018-04-25 | Steganography detection method for multi-embedding-rate encrypted image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108805161B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110059754B (en) * | 2019-04-22 | 2019-12-24 | 厦门大学 | Batch data steganography method, terminal equipment and storage medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103258123A (en) * | 2013-04-25 | 2013-08-21 | 中国科学院信息工程研究所 | Steganalysis method based on blindness of steganalysis systems |
| CN106203492A (en) * | 2016-06-30 | 2016-12-07 | 中国科学院计算技术研究所 | The system and method that a kind of image latent writing is analyzed |
| CN106504177A (en) * | 2016-09-30 | 2017-03-15 | 广东技术师范学院 | A kind of low embedded rate compressed sensing general steganalysis method of coloured image |
| CN107273916A (en) * | 2017-05-22 | 2017-10-20 | 上海大学 | The unknown Information Hiding ＆ Detecting method of steganographic algorithm |
-
2018
- 2018-04-25 CN CN201810375675.2A patent/CN108805161B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103258123A (en) * | 2013-04-25 | 2013-08-21 | 中国科学院信息工程研究所 | Steganalysis method based on blindness of steganalysis systems |
| CN106203492A (en) * | 2016-06-30 | 2016-12-07 | 中国科学院计算技术研究所 | The system and method that a kind of image latent writing is analyzed |
| CN106504177A (en) * | 2016-09-30 | 2017-03-15 | 广东技术师范学院 | A kind of low embedded rate compressed sensing general steganalysis method of coloured image |
| CN107273916A (en) * | 2017-05-22 | 2017-10-20 | 上海大学 | The unknown Information Hiding ＆ Detecting method of steganographic algorithm |
Non-Patent Citations (8)
| Title |
|---|
| "An Adaptive Ensemble Classifier for Steganalysis Based on Dynamic Weighted Fusion";Xikai Xu et al.;《Proceeding of the 2015 International Conference on Communications,Signal Processing,and Systems》;20160623;全文 * |
| "JPEG Steganalysis With High-Dimensional Features and Bayesian Ensemble Classifier";Fengyong Li et al.;《IEEE Signal Processing Letters》;20130331;第20卷(第3期);全文 * |
| "Mixing high-dimensional features for JPEG steganalysis with ensemble classifer";Bin Chen et al.;《SIViP》;20141231;全文 * |
| "基于不确定性推理的JPEG图像通用隐藏信息检测技术";朱婷婷 等;《电子学报》;20130228(第2期);全文 * |
| "基于多分类器融合的位置嵌入率图像隐写分析方法";万宝吉 等;《数据采集与处理》;20140930;第29卷(第5期);全文 * |
| "基于层次修剪和加权投票的图像隐写分析";曹镇 等;《计算机工程与设计》;20170930;第38卷(第9期);全文 * |
| "基于高维特征正则化的JPEG图像隐写分析";孙物一 等;《应用科学学报》;20160930;第34卷(第5期);全文 * |
| "深度学习的图像隐写分析方法研究";钱银龙;《中国博士学位论文全文数据库 信息科技辑》;20180215(第02期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108805161A (en) | 2018-11-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Ciftci et al. | How do the hearts of deep fakes beat? deep fake source detection via interpreting residuals with biological signals | |
| Zhao et al. | Towards more accurate iris recognition using deeply learned spatially corresponding features | |
| Hosler et al. | A video camera model identification system using deep learning and fusion | |
| CN111160313B (en) | Face representation attack detection method based on LBP-VAE anomaly detection model | |
| Taimori et al. | A novel forensic image analysis tool for discovering double JPEG compression clues | |
| Zois et al. | A comprehensive study of sparse representation techniques for offline signature verification | |
| Kiros et al. | Deep representations and codes for image auto-annotation | |
| Wu et al. | Steganalysis via deep residual network | |
| Mayer et al. | Open set video camera model verification | |
| Thing et al. | An improved double compression detection method for JPEG image forensics | |
| Goh et al. | A hybrid evolutionary algorithm for feature and ensemble selection in image tampering detection | |
| CN107273916A (en) | The unknown Information Hiding ＆ Detecting method of steganographic algorithm | |
| Castillo Camacho et al. | A simple and effective initialization of CNN for forensics of image processing operations | |
| CN108805161B (en) | Steganography detection method for multi-embedding-rate encrypted image | |
| Pevný et al. | Determining the stego algorithm for JPEG images | |
| González-Soler et al. | Fisher vector encoding of dense-bsif features for unknown face presentation attack detection | |
| Deshpande et al. | Neural network based block-level detection of same quality factor double JPEG compression | |
| Cattaneo et al. | Experimental evaluation of an algorithm for the detection of tampered JPEG images | |
| Zhao et al. | Source camera identification based on coupling coding and adaptive filter | |
| Huang et al. | A method for identifying origin of digital images using a convolutional neural network | |
| CN111178204B (en) | Video data editing and identifying method and device, intelligent terminal and storage medium | |
| Chen et al. | A study on the photo response non-uniformity noise pattern based image forensics in real-world applications | |
| Pondenkandath et al. | Leveraging random label memorization for unsupervised pre-training | |
| Boroumand et al. | Nonlinear feature normalization in steganalysis | |
| Kauba et al. | Identifying the origin of iris images based on fusion of local image descriptors and PRNU based techniques |
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 |