CC BY
9
UDC 621.39
Investigation of Digital Image Preprocessing Methods Influence on the Accuracy of Stego
Images Detection
Progonov D. O.
National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute", Kyiv, Ukraine
E-mail: progonov&gmaiL com
The feature of modern methods of detecting unauthorized transmission of confidential data in communication systems is widespread usage of pre-processing methods for transmitted files, such as digital images. The purpose of these methods is to detect weak changes of cover image's statistical parameters cauased by message hiding. A significant number of these methods are based on usage of ensembles of high-pass filters, which allows to ensure high accuracy of detection of steganograms (more than 95%) formed according to known st.eganographic methods. However, a significant limitation of the practical application of these methods is high computational complexity of ensemble forming procedure that minimizes the detection error of stego images. This makes it impossible to quickly reconfigure stegdetect.ors to detect stego images formed according to a priori unknown embedding methods. Therefore, it is of special interest to develop fast methods for image pre-processing, which can reliably detect weak changes of cover's statistical parameters under limited a priori information about used st.eganographic method. The work is devoted to the study of the achievable accuracy of the stedetect.or with variations type and parameters of digital images pre-processing methods. According to the results of the study, the optimal methods of pre-processing image to minimize
9
the error of stego images detection compared to modern pre-processing methods, even in the most difficult case of low payload of cover image (less than 10%) and limited a priori data about used embedding method. It is revealed that usage of special types of image pre-processing methods, namely denoising autoencoders, allows to bring the accuracy of a stegdetector closer to the proposed estimations of achievable accuracy of stegodetect.ors.
Keywords: st.eganalysis: stego image preprocessing methods: digital images DOI: 10.20535/RADAP. 2022.89.54-60
Acronyms
ANN Artificial Neural Network CI Cover Image
CNN Convolutional Neural Network DAE Denoising Antoencoder DI Digital Image HPF High-Pass Filter PPM Preprocessing Methods SD Stegdetector SM Steganographic Method
Introduction
Ensuring of reliable protection of sensitive data from unauthorized transmission (leakage) is topical
task today. Of special interest are methods related to revealing and counteraction of hidden (concealed) data transmission with nsage of multimedia data, snch as Digital Image (DI) [1]. These methods are based on recent advances of digital media steganography. namely adaptive message hiding in DI. that makes revealing of formed stego images a non-trivial task.
The state-of-the-art approach to detect of stego images formed by novel Steganographic Method (SM) is based on analysis differences between the statistical, spectral, and structural parameters of initial (cover) and processed images [1]. This makes possible nsage of multidimensional vector classification methods to reveal weak alterations of Cover Image (CI) parameters caused by stegodata embedding. However, the detection accuracy of novel Stegdetector (SD) hardly depends on the availability of prior information about used embedding methods that limits nsage of SD in real cases where snch information is limited or even absent.
One of proposed approaches to overcome mentioned limitation of modern SD is based on applying of Preprocessing Methods (PPM) to analyzed DI. The pre-processing methods are aimed at revealing and emphasizing weak alterations of CI pixels brightness caused by message hiding. Variety of proposed PPM makes possible fine-tuning of a SD to achieve high detection accuracy (more than 90%) for wide range of modern embedding methods. However, there is no information in the literature about the achievable detection accuracy for stego images by using PPM of specific type. This makes it difficult to choose effective pre-processing methods of DI to minimize the error of stego images detection especially for unknown embedding methods (zero-day problem). Tims, the topical task is estimation of achievable accuracy of stego images detection in case of limited or even absent prior information about used SM.
1 Related works
The considerable amount of proposed PPM for modern stegdetectors is based on usage an ensemble of High-Pass Filter (HPF) [1]. Applying of such filters allows effectively suppression of cover image content and further extraction of high-frequency components used for message hiding. Despite high accuracy of stego images detection (more than 95%) by usage of such approach to DI preprocessing, practical use of the corresponding SD is limited. This is caused by necessity of time-consuming and resource-intensive optimization of ensemble to minimize the stego images detection error.
To overcome this limitation, it was proposed to use Artificial Neural Network (ANN), in particular Convolutional Neural Network (CNN) and Denoising Autoencoder (DAE). The feature of ANN-based stegdetectors is ability to adapt neural networks parameters during training (backpropagation procedure) to minimize stego images detection error Pe [ ]• This allows selecting the optimal parameters of input (convolutional) layers by criterion of error PE minimization during backpropagation procedure instead of painstaking selection of HPF ensembles elements. In contrast to CNN. the DAE is based on projection of multidimensional vectors (related to statistical parameters of analyzed images) to a space of smaller dimensions while preserving their relative positions [2]. This makes it possible to reduce of Pe value on new datasets of in comparison with the widespread stegdetectors [5]. Despite outstanding detection accuracy of SD based on considered CNN and DAE architectures, tuning of such stegdetectors remain resource-intensive procedure. Also, trained stegdetectors may remain vulnerable to change of used image dataset (domain mismatch problem) due to usage of shallow neural networks caused by computation resource constraints.
Therefore, of special interest are development of advanced PPM methods that provide high detection accuracy of CI alteration caused by message hiding by preserving fixed (low) computation complexity. Despite variety of proposed PPM methods, selection of appropriate image pre-processing method that minimizes Pe under specified and used image dataset requires exhaustive search among known PPM methods [6]. This is caused by absence of theoretical foundations for the selection of optimal PPM based on the criterion of minimizing the value of the classification error of stego images Pe- Solving of mentioned problem requires estimation achievable detection accuracy depends on available information about used embedding method and digital image dataset. However, information presented in the literature covers only specific cases, such as standard image datasets or widespread SM. Thus, the work is aimed at filling this gap by analysis of achievable Pe values level by-using modern PPM depending on the available priori information about used SM.
2 Task and challenges
The purpose of the work is estimation of achievable detection accuracy of stego images by varying the type of image pre-processing methods and limitations of priori data about SM. The results of research allows developing the methodology for determining the optimal PPM method by criterion of minimizing stego images detection error for modern steganographic methods.
3 Analysis of the achievable detection accuracy for moderd stegdetectors
The detection accuracy of the pre-trained SD can be represented in the following form, depending on the selected PPM and methods for DI features extraction:
PE(I) = /(F, S, K, I, X, y), (1)
where Pe is stego images detection error; I is F
S
K
parameters of the analyzed image with corresponding characteristics of the cover and stego images; X, y are sets of cover and stego images respectively that are available during SD training. K
of PPM to analyzed image in order to detect weak changes of its statistical parameters cause by message embedding. Currently, the ensembles of HPF are widely used as such operator [3.6]. At the next stage, the S
components of processed analyzed image for estimation their statistical, spectral, and structural parameters [1]. However, ensuring high accuracy of stego images 95%
number of high-pass filters (e.g. 12.870 parameters for the PSRM model. 34.671 for the maxSRM model), which complicates tuning of the SD. Finally, extracted
K
in eq. (1). The ensemble classifiers, such as random Forests, are widely used for performing this task to minimize the value of the error Pe while preserving low computational complexity of the SD tuning [8].
Despite the emergence of advanced methods for analysis and classification of DI statistical parameters for design of high-precision SD. the selection of optimal to minimize the Pe values currently is not given enough attention by researches. The common way to compare the effectiveness of PPM methods is based on investigation of shift for multidimensional vectors (statistical parameters) of cover and stego images caused by image pre-processing [7]. A schematic representation of the changes in the mutual position of these feature vectors for both cover and stego images after the applying of PPM is shown at Fig. 1.
Fig. 1. The influence of digital image preprocessing methods on the mutual positions of vectors (statistical characteristics) of cover and stego images. According to materials [7].
Message embedding to the CI leads to shift of the corresponding feature vectors (statistical parameters) of the cover images F(c) % the amount me to the new position F(s). In this case, the spread of the values of the vectors F(c) is denoted as Mc, while the corresponding spread of the vectors F(s) for formed stego images is equal to Me. Applying of methods leads to a shift of the vectors F(c) and F(s) by the corresponding values mrc mid mrs to the new positions Fr(c) and Fr(s). Reciprocally, the value of the ''spread" of positions for feature vectors related to processed cover and stego images equals to Mrc and Mg respectively.
Applying of PPM to the studied DI leads to corresponding changes of mutual positions for clusters of extracted feature vectors of cover and stego images. In the work [7], the following classification of DI preprocessing methods was proposed:
1. Parallel reference application of PPM leads to the similar changes of characteristics for cover Fr(c) and stego Fr (s) images;
2. Eraser application of these methods leads to a decrease in the distance between Fr (c) and Fr (s) compared to the distance between the vectors F(c) Mid F(s);
3. Divergent reference aimed at increasing the distance between clusters of vectors Fr (c) and Fr (s);
4. Cover estimate is aimed at evaluating the statistical features of the CI based on the available (noisy) images;
5. Stego estimate is aimed at detecting distortions of the CI caused by hiding the recorded data.
Let us note that usage of PPM methods related to parallel reference or eraser cases is limited today. This is caused by negligible impact on distance between the clusters of vectors Fr(c^d Fr (s) and, respectively, detection accuracy. On the other hand, divergent reference methods allow considereably improving detection accuracy by strengthening the differences between the cover and stego images. However, practical usage of such methods requires precise assessment relative position of feature vectors clusters Fr (c) and Fr(s). This needs usage of prior information about features of used embedding method that may be limited or even absent in real cases.
As a result, a significant number of modern PPM are aimed at estimating the parameters of the CI based on the available (noisy) data (cover estimate methods), or detecting weak changes of CI parameters caused by message hiding (stego estimate methods) [9, 10]. However, there is limited information about achievable accuracy of SD by using these types of PPM.
Let us consider the case of applying idealized PPM methods related to the cover (Ffd^al(•)) and stego (Festimation cases:
F?deai(X, Y): X ^ X, Y (Af)
VAS>0
X,
F SE
i de a I
(X, Y) : X
VAS>0
■> Y (Af),
Y (Af)
A^=c on st
-+Y (Af) , (3)
X, Y
Af is the cover image payload.
Let us denote the statistical parameters for original and processed images by the corresponding vectors Fnc and F caiib. Then, applying of considered methods in eqs. (2) and (3) leads to the case when either
features of cover, or stego images are preserved after transformation while counterpart features are changed considerably. For example, the elements of F c aiib vector will be the same for cover (eq. (2)) or stego (eq. (3)) images, which will significantly reduce the accuracy of the . In this case, vectors F cc = {Fnc; Fcaiib} will include the features of original and processed images, which is widely used in modern SD. On the other hand, the magnitude of vectors F of = Fnc — Fcalib will be proportional to the CI features distortions caused by stego data embedding.
Then, accuracy of estimation of CI distortions caused by message hiding by using vectors F of depends on the accuracy of restoration cover image from current (noisy) ones (by using the operator FcJ^a,(X, Y)), or stego one (by using the operator Ff^t(X, Y)). Respectively, usage of "idealized" operators (eqs. (2) and (3)) will allow minimizing the stego image detection error for fixed methods for images features estimation and classification. Thus, applying of vectors F of makes possible estimation of achievable level of Pe error by using standard methods for DI feature extraction and classification.
4 R,GSults
arbitrary CI). This allows studying the influence of prior informationabout used embedding methods on the detection accuracy of SD [14].
The standard statistical model SPAM [15] was used to estimate parameters of analyzed DI, while the assignment of the analyzed image to classes of cover or stego images was carried out using random Forest classifier [8]. Also, the cases of usage cover rich model maxSRMd2 [6], as well as advanced artificial neural networks GB-Ras [4] and ASSAF [5] were considered. The maxSRMd2 model is based on usage HPF ensemble for DI pre-processing, and the further applying of Markov chain models to estimate correlation of adjacent pixels brightness [6]. The GB-Ras convolutional neural network allows ensuring high detection accuracy for wide range of SM due to usage of specialized convolutional methods in the input layers, namely depthwise separable convolution [4]. In contrast to the GB-Ras network, the hybrid ASSAF network is based on usage of DAE to estimate parameters of CI using the available (noisy) data, and further analysis differences between these parameters. This makes possible to ensure high detection accuracy for ASSAF network even under limited prior information about used SM [5].
4.1 Experimental setup
Performance analysis of achievable accuracy of stego images detection by usage of proposed DI preprocessing methods (eqs. (2) and (3)) was carried out on the standard image package ALASKA [11]. The case of applying modern steganographic methods HUGO [12] and MiPOD [13] for the message embedding into CI was considered. These methods are based on presentation of embedding process as solving of corresponding optimization problem minimizing changes of CI statistical parameters during embedding a message with fixed bitlength. The CI payload values were varied in the following range — 3% 5% 10% 20%, 30% 40% 50%.
The analysis of detection accuracy was perfomed according to standard cross-validation procedure. The used images dataset was divided into training Strain (70%) and Stest testing (30%) subsets. The partitioning was repeated 10 times to obtain the average Pe values. Images from the training and testing sets were scaled to the identical resolution of 512 x 512 pixels.
The tuning of SD was done by varying the ratio of cover and stego images pairs (K°L) in the Strain set:
TsOL _ \(X Y) : (Xi, Yi), * e Strain 1 = -
\Strain 1
When conducting research, the values of this indi-0%
to the encoder and can only use available stego images) 100%
4.2 Experiments
The study of the detection accuracy by variation of PPM types was carried out in several stages. At the first stage, the analysis of the achievable accuracy of the SD was done by using the considered pre-processing methods eqs. (2) and (3). The dependence of stego images detection error Pe on cover image payload for stego images formed according to the HUGO steganographic method, when using the methods eqs. (2) and (3) are presented in Fig. 2.
x 100%. (4)
Fig. 2. Dependencies of stego images detection error Pe on cover image payload for stego images formed according to the HUGO steganographic method, when using preprocessing methods eqs. (2) and (3) for the ALASKA package and value of K°°L parameter: (a) — KOL = 100%; (b) KOL eU(0; 100); (c) KOL = 0%.
The decrease of the K°°L value from 100% to 0% leads to corresponding increase of Pe by 7% (Fig. ). This is caused by a gradual reducing the ratio of cover-to-stego pairs in the Strain set.
Applying of F caiib vectors leads to slight changes of PE values by varying of cover image payload (Fig. 2). This is caused by the fact that the statistical characteristics of the DI after applying of preprocessing methods eqs. (2) and (3) will coincide with the corresponding characteristics of the cover or stego images. As a result, the detection accuracy of SD will be determined only by the parameters of used feature extraction and classification methods.
On the other hand, applying of the considered methods eqs. (2) and (3) leads to the proportional to changes of elements both for both F of vectors and statistical parameters of CI caused by message hiding. As a result, the detection accuracy significantly increases (Fig. 2) especially for the case of low cover image payload (less than 10%), where the efficiency of modern SD is usually low.
The use of vectors F cc during tuning also leads to decrease of Pe values compared to the case of using F„ c vectors (Fig. ). This is explained by the F c c
cs of both initial and processed images, which leads to a doubling of the number of data elements of the vectors and corresponding increasing of complexity of SD adjustment.
For comparison, Fig. 3 shows the dependence of the values of the classification error Pe on cover image payload for stego images formed according to the MiPOD steganographic method, when using image preprocessing methods eqs. (2) and (3).
MiPOD, compared to the case of using the HUGO embedding method.
At the second stage of research, the comparative analysis of detection accuracy for stego images formed according to the considered embedding methods was performed. The case of using both preprocessing methods eqs. (2) and (3) and modern SD was considered. The value of the detection error Pe for stego images, formed according to the HUGO and MiPOD steganographic methods, for SD based on usage of pre-processing methods eqs. (2) and (3), as well as statistical models SPAM [15] and maxSRMd2 [6], artificial neural networks GB-Ras [4] and ASSAF [5] are given in table 1.
Table 1 The value of detection error Pe for stegodetectors based on usage of proposed image preprocessing methods FcdEal(X, Y^d Ff^l(X, Y), statistical model maxSRMd2 and artificial neural networks GB-RAS and ASSAF at K°L = 0% for the ALASKA image dataset
Type of stegdetector Cover image payload, %
5% 10% 30% 50%
HUGO embedding method
F df 6.02 3.30 1.01 0.67
F cc 23.45 13.43 5.06 3.07
SPAM 56.46 56.50 52.29 47.76
maxSRMd2 47.33 44.44 35.04 28.59
GB-Ras 49.51 48.73 46.27 43.67
ASSAF 11.99 11.70 10.56 9.63
MiPOD embedding method
F df 3.25 2.07 0.79 0.49
F cc 13.11 8.01 3.37 2.28
SPAM 55.97 55.88 51.75 45.80
maxSRMd2 48.39 45.84 36.52 29.50
GB-Ras 49.32 48.57 46.16 44.20
ASSAF 12.18 12.09 12.03 12.14
Fig. 3. Dependencies of the error values PE on cover image payload for stego images formed according to the MiPOD steganographic method, when using preprocessing methods (2)-(3) for the ALASKA package and the value of the parameter K°L: (a) — K°L = 100%; (b) - K°L e U(0; 100); (c) - K°L =0%.
The obtained results for MiPOD embedding method (Fig. 3) confirm the previously obtained results for the HUGO method (Fig. 2), namely significant increase of detection accuracy by using image preprocessing methods eqs. (2) and (3). It is worth to note the reduction of PE values on 5% — 7% % using F of F c c
to the corresponding values for the HUGO method Conclusion (Fig. 2). This indicates about greater changes of adjacent pixels brightness correlation by data embedding, according to the novel steganographic method
The use of considered image pre-processing methods eqs. (2) and (3) makes it possible to reduce the value of detection error PE up to 9 times in comparison with modern SD based on cover rich models and ANN. At the same time, the revealed decreasing of PE values is also preserved even for low cover image payload (less than 10%), which is of particular interest when performing the stegoanalysis of advanced embedding methods.
It is worth to note that usage of special types of PPM, namely denoising autoencoder, allows approaching the achievable accuracy of stego images
detection when using Fcd^a l(X, Y) and Ftfd^ai(X, Y). This proves the prospects of using DAE for preprocessing of analyzed images to increase detection accuracy of SD.
The estimation of achievable detection accuracy of stego images by varying the type of image pre-
processing methods and limitations of priori data about SM was performed. The optimal methods of preprocessing of DI based on the criterion of minimizing the stego images detection error are determined. It was revealed that applying of image-preprocessing methods aimed at restoration of cover image based on the available (noisy) data and the isolation of DI changes caused by stegodata embedding allows approaching to estimated achievable detection accuracy. It was shown that application of these methods allows reducing the detection error value for stego images formed according to modern HUGO and MiPOD steganographic methods up to 9 times, even for the case of low cover image payioad (less than 10%) and the limitations of a priori data regarding the characteristics of embedding methods.
It has been established that the use of specialized image preprocessing methods, namely denoising autoencoders, allows ensuring detection accuracy, which is comparable to the use of the proposed preprocessing methods. Thus, research effectiveness of stegdetectors based on DAE for processing real DI, which are characterized by high variability of statistical parameters, is of further interest.
References
[1] Hassaballah M. ("2020) Digital Media Steganography: Principles, Algorithms, and Advances, Academic Press. 386 p.. ISBN: 978-0128194386.
[2] Goodfellow 1.. Bengio Y. and Courville A. (2016) Deep Learning, The MIT Press. 800 p.. ISBN: 978-0262035613.
[3] Boroumand M. Chen M. and Fridrich .1. (2019). Deep Residual Network for Steganalysis of Digital Images. IEEE 'transactions on Information Forensics and Security, Vol 14. No. 5. p. 1181-1193. DOl: 10.1109/T1FS.2018.2871749.
[4] Tabares-Soto R.. Arteaga-Arteaga H. B.. Bravo-Ortiz M. A.. Mora-Rubio A.. Arias-Garzon D.. Alzate-Grisales .1. A.. Burbano-.Iacome A. B.. Orozco-Arias S.. Isaza G. and Ramos-Pollan R. (2021). GBRAS-Net: A Convoluti-onal Neural Network Architecture for Spatial Image Steganalysis. IEEE Access, Vol 9. p. 14340-14350. DOl: 10.1109/ACCESS.2021.3052494.
[51 Cohen A.. Cohen A. and Nissim N. (2020). ASSAF: Advanced and Slim StegAnalysis Detection Framework for JPEG images based on deep «involutional denoising autoencoder and Siamese networks. Neural Networks, Vol 131. p. 64-77. DOl: 10.1016/j.neunet.2020.07.022.
[6] Denemark T.. Sedihi V., Holub V., Cogranne R. and Fridrich .1. (2014). Selection-channel-aware rich model for Steganalysis of digital images. International Workshop on Information Forensics and Security, Atlanta. USA. DOl: 10.1109/W1FS.2014.7084302.
[7] Kodovsky .1. and Fridrich .1. (2009) Calibration revisited. 11th ACM workshop on Multimedia and security, ACM. p. 63-74. DOl: 10.1145/1597817.1597830.
[8] Kodovsky .1.. Fridrich .1. and Holub V. (2011) Ensemble Classiliers for Steganalysis of Digital Media. IEEE 'transactions on Information Forensics and Security, Vol. 7. Issue 2. p. 432-444. DOl: 10.1109/T1FS.2011.2175919.
[9] Progonov D. (2021) Statistical stegdetectors performance by message re-embedding. 'theoretical and Applied Cybersecurity, Vol.3. No. 1. p. 5-14. DOl: 10.20535/tacs.2664-29132021.1.251291.
[10] Progonov D.O. (2021) Eifectiveness of stego images pro-noising with fractional noise for digital image steganalysis. Applied Aspects of Information 'technology, Vol. 4. issue 3. p. 261-270. DOl: 10.15276/aait.03.2021.5.
[Ill Cogranne R.. Giboulot Q. and P. Bas. (2019) The ALASKA Steganalysis Challenge: A First Step Towards Steganalysis. ACM Workshop on Information Hiding and Multimedia Security, New York. USA. p. 125-137. DOl: 10.1145/3335203.3335726.
[12] Filler T. and Fridrich .1. (2010) Gibbs Construction in Steganography. IEEE 'transactions on Information Forensics and Security, Vol. 5. Issue 4. p. 705-720. DOl: 10.1109 / TIF S.2010.2077629.
[131 Sedighi v.. Cogranne R. and Fridrich .1. (2015) Content-Adaptive Steganography by Minimizing Statistical Detectability. IEEE 'transactions on Information Forensics and Security, Vol. 11. Issue 2. DOl: 10.1109/T1-FS.2015.2486744.
[14] Progonov D. (2020) Performance of Statistical Stegdetectors in Case of Small Number of Stego Images in Training Set. IEEE International Scientific-Practical Conference "Problems of lnfocom.rn.unicati.ons Science and Technology", Kharkiv. Ukraine. DOl: 10.1109/P1CST51311.2020.9467901.
[15] Pevny Т.. Bas P. and Fridrich .1. (2010) Steganalysis by Subtractive Pixel Adjacency Matrix. IEEE 'transactions on Information Forensics and Security, Vol. 5. Issue 2. p. 215-224. DOl: 10.1109/T1FS.2010.2045842.
Дослщження впливу метод!в поперед-Hboi" обробки цифрових зображень на точшсть виявлення стеганограм
Прогонов Д. О.
Особлишстю сучаспих метод!в виявлеппя песапкцю-повапо! (приховапо!) передач! копф1депццших дапих в шформацшпо-комушкацццшх мережах е широке вико-ристаппя метод!в попередпьо! обробки досшджувапих фай.шв. зокрема цифрових зображень. Метою дапих метод!в е детектуваппя слабких змш статистичпих па-раметр!в зображеппя-коптейперу. обумовлепих прихо-ваппям пов1домлепь. Зпачпа шльшсть дапих метод!в засповапа па використапш апсамбл1в ф!льтр1в висо-ких частот, що дозволяв забезпечити високу точшсть виявлення стеганограм (бшыне 95%), сформованих 3ri-дпо в!домих стегапограф1чних метод!в. Проте вагомим обмежепиям практичного застосуваппя запроиопованих метод!в попередпьо! обробки цифрових зображень е ви-сока обчислювальпа складшсть ироцедури формуваш1я дапих апсамбл!в для мппм1зацп помилки виявлеппя стеганограм. Це упеможливлюе швндке перепалаштува-ппя стегодетектор!в для виявлешш стеганограм. сформованих зпдпо anpiopno пев!домих стегапограф1чних метод!в. Тому стаповить 1птерес розробка швидких ме-тод!в попередпьо! обробки досл1джувапих зображень. здаташх падишо виявляти слабк! змшп статистичних па-раметр!в зображмшя-коптейперу в умовах обмежепост! anpiopmix дапих щодо використапого стегапограф1чпого
методу. Робота присвячена доондженню досяжно! то-чноста роботи стегодетектору при вар!аци метод!в попередньо! обробки цифрових зображень. За результатами доондження визначено оптимальш методи попередньо! обробки зображень для мш1м1зацп помилки виявлен-ия стегаиограм. Даш методи дозволяють суттево (до 9 раз!в) зменшити иомилку класифшаци стегаиограм у пор!внянш з сучасними методами попередньо! обробки зображень, нав!ть у найбшын складному випадку
слабкого заповнення зображення-контейнеру стегодани-ми (менше 10%) та обмеженоста апрюрнпх даних щодо використаного стеганограф!чного методу. Виявлено, що використання спегцал1зованпх метод!в обробки зображень, а саме знешумлюючих автоенкодер!в, дозволяв наблизити точшсть роботи стегодетектор!в до отрима-них оцшок досяжно! точноста роботи стегодетектор!в.
Ключовг слова: стегоанал!з; методи попередньо! обробки стегаиограм; цифров! зображення
