Pixels-off: Data-augmentation Complementary Solution for Deep-learning Steganalysis
29
0
0
Texte intégral
(2) Introduction and background. Pixels-off technique. Conclusions. Outline. Introduction and background Pixels-off technique Conclusion. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 1/22.
(3) Introduction and background. Pixels-off technique. Conclusions. Steganography & Steganalysis secret key Alice. Bob. message stego. Emb. cover Eve. Ext. message. cover or stego. Text. Steganography vs. Steganalysis Steganography: the practice of concealing a secret message within a digital support. Steganalysis: the analysis of a cover material to identify the presence of hidden information. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 2/22.
(4) Introduction and background. Pixels-off technique. Conclusions. The chronology of steganalysis DL models evolution. 2015 jan. 2016 june. jan. S. Tan and B. Li EI Asia-Pacific’2014 Stacked auto-encoder. Qian et al. [Tan] EI’2015 1ere description CNN. 2017 june. IH&MMSec ICIP. EI. Pibre et al. [Chaumont] EI’2016 Same keys. EI. IH&MMSec. june. ICIP’2016 Transfer learning. Ye et al. [Yi] TIFS’2017 2nd reference network Ye-Net. Chen et al. [Fridrich] IH&MMSec’2017 JPEG : ad hoc topology. dec. jan. june. dec. IH&MMSec. EI Yedroudj et al. [Chaumont] Data-base augmentation. Qian et al. [Tan] IEEE Sig. Proc. Letters 1st reference CNN (close to AlexNet) Xu-Net. dec jan. Fuji Tsang and Fridrich EI’2018 Images of arbitrary size Chen et al. [Fridrich] EI’2018 Quantitative Steganalysis. IH&MMSec’2016 Ensemble version. ..... 2019. 2018 june. Xu et al. [Shi]. Xu et al. [Shi]. SPATIAL. jan. Zhang et al.. Yedroudj et al. [Chaumont] IEEE ICASSP Another reference CNN Yedroudj-Net. TIFS’2019. Deng et al. H&MMSec’2019. 3 improvements on Yedroudj-Net (Zhu-Net). 3 improvements on CovPool-Net. Li et al.. Zeng et al. [Li & Huang]. IEEE Sig. Proc Letters 2018 Boroumand et al. [Fridrich] ReST-Net : Combinaison TIFS’2018 of 3 CNNs For Spatial and JPEG Another reference CNN SRNet. ReST-Net for color images (WIReST-Net). Xu (Xu-Net-Jpeg) IH&MMSec’2017 JPEG : close to Res-Net. JPEG. Mehdi Yedroudj. Huang et al.. Zeng et al. [Huang]. Zeng et al. [Huang]. EI’2017 JPEG : Large Scale. TIFS’2018 JPEG : Large Scale. H&MMSec’2019. Low complexity for JPG steganalysis LowComplexity -Net. Pixels-off: Data-augmentation Complementary Solution. 3/22.
(5) Introduction and background. Pixels-off technique. Conclusions. The evolution of steganalysis models depth over the last 5 years. Series 1 linear regression line. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 4/22.
(6) Introduction and background. Pixels-off technique. Conclusions. NN’s performance in terms of depth and amount of data. Inspired from [1] [1] Deep neural networks are able to learn from massive amounts of data — adapted from ‘AI is the New Electricity’ (Andrew Ng). Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 5/22.
(7) Introduction and background. Pixels-off technique. Conclusions. Steganalysis models performance in terms of depth and amount of data Test protocol I Stéganographie: WOW I Payload: 0.2bpp. Xu-Net Yedroudj-Net Ye-Net SRNet. BOSS (4000) 32.4 % 27.8 % 33.1 % 32.5 %. BOSS+BOWS2 (14000) 30.3 % 23.7 % 26.1 % 24.1 %. BOSS+BOWS2+VA (112000) 30.5 % 20.8 % 22.2 % 19.0 %. [2] M. Yedroudj, M. Chaumont, F. Comby YEDROUDJ-NET:An efficient CNN for spatial steganalysis. (ICASSP), 2018. Xu-Net:5conv Mehdi Yedroudj. Ye-Net:8conv. Yedroudj-Net:5conv. SRNet:14. Pixels-off: Data-augmentation Complementary Solution. conv. 6/22.
(8) Introduction and background. Pixels-off technique. Conclusions. Existing solutions for data enrichment Given a target database: For the learning of the NN: I Apply straightforward virtual data augmentation in either online or offline manner (flip & rotation → ×8), I Use other database similar to the target database (e.g. BOSS+BOWS2), I Use similar cameras to capture new images, and reproduce the same development than the target database, I Apply similar developments to those in the target database on the original RAW images. [3] M. Yedroudj, M. Chaumont, F. Comby How to augment a small learning set for improving the performances of a CNN-based steganalyz. (EI), 2018. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 7/22.
(9) Introduction and background. Pixels-off technique. Conclusions. Problematic: Real life case scenario Limitation of existing data augmentation solutions I Due to storage limitations, RAW images are not usually available, besides not easy to reproduce the same development: I Apply similar developments to those in the target database on the original RAW images.. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 8/22.
(10) Introduction and background. Pixels-off technique. Conclusions. Problematic: Real life case scenario Limitation of existing data augmentation solutions I Due to storage limitations, RAW images are not usually available, besides not easy to reproduce the same development: I Apply similar developments to those in the target database on the original RAW images.. I Definition of database ’resemblance’ is not yet well established (no objective measurement): I Use other database similar to the target database (i.g. BOSS+BOWS2) ,. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 8/22.
(11) Introduction and background. Pixels-off technique. Conclusions. Problematic: Real life case scenario Limitation of existing data augmentation solutions I Due to storage limitations, RAW images are not usually available, besides not easy to reproduce the same development: I Apply similar developments to those in the target database on the original RAW images.. I Definition of database ’resemblance’ is not yet well established (no objective measurement): I Use other database similar to the target database (i.g. BOSS+BOWS2) ,. I The enormous variety of existing digital cameras: I Use similar cameras to capture new images, and reproduce the same development than the target database,. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 8/22.
(12) Introduction and background. Pixels-off technique. Conclusions. Problematic: Real life case scenario Limitation of existing data augmentation solutions I Due to storage limitations, RAW images are not usually available, besides not easy to reproduce the same development: I Apply similar developments to those in the target database on the original RAW images.. I Definition of database ’resemblance’ is not yet well established (no objective measurement): I Use other database similar to the target database (i.g. BOSS+BOWS2) ,. I The enormous variety of existing digital cameras: I Use similar cameras to capture new images, and reproduce the same development than the target database,. I Very small amount of data to start with (10,100 images): I Apply straightforward virtual data augmentation in either online or offline manner (flip & rotation → ×8) Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 8/22.
(13) Introduction and background. Pixels-off technique. Conclusions. Proposed approach. Pixels-off technique I A new way to enrich a database in order to improve the CNN-based steganalysis performance, I An efficient, generic approach which is usable in conjunction with other data-enrichment approaches, I It can be used to build a ”Side-Channel-Aware database” (SCA-database). Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 9/22.
(14) Introduction and background. Pixels-off technique. Conclusions. Global flowchart of the pixel-off technique. Phase1: covers preparation. Phase2: stegos generation Phase3: training with data augmentation. covers. Pixels-off. covers. covers. +. . . . . . .. ..covers .. . . ...V.1 . ... covers. . . . . . .. ..covers .. . . ...V.1 . ... Data embedding. stegos. . . . . . .. ..stegos .. . . ...V.1 . ... covers covers. stegos. . . . .. covers .. . . . V.1 . ... stegos. . . . .. stegos .. . . . V.1 . ... Model training. Covers+pixels off. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 10/22.
(15) Introduction and background. Pixels-off technique. Conclusions. Phase1 : covers preparation Phase1: covers preparation. Phase2 : Stego generation. tra covers. covers. Pixels-off. covers covers. +. . . . . . .. ..covers .. . . ...V.1 . ... Data embedding. . . . . . .. ..covers .. . . ...V.1 . ... stegos. . . . . . .. stegos .. .. . . ...V.1 . ... Covers+pixels off. Protocol I Set the value of ”P”, the number of pixels to switch off, I Generate a pixels-off version of cover images, I A new set of covers is produced. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 11/22.
(16) Introduction and background. Pixels-off technique. Conclusions. Phase1 : covers preparation Phase1: covers preparation. Phase2 : Stego generation. tra. . . . . . .. ..covers .. . . ...V.1 . ... covers. covers. Pixels-off. Data embedding P: number of zeros pixels. Pixels-off. covers covers. +. . . . . . .. ..covers .. . . ...V.1 . ... stegos. . . . . . .. stegos .. .. . . ...V.1 . ... Covers+pixels off. Protocol I Set the value of ”P”, the number of pixels to switch off, I Generate a pixels-off version of cover images, I A new set of covers is produced. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 11/22.
(17) Introduction and background. Pixels-off technique. Conclusions. Global flowchart of the pixel-off technique. Phase1: covers preparation. Phase2: stegos generation Phase3: training with data augmentation. covers. Pixels-off. covers. covers. +. . . . . . .. ..covers .. . . ...V.1 . ... covers. . . . . . .. ..covers .. . . ...V.1 . ... Data embedding. stegos. . . . . . .. ..stegos .. . . ...V.1 . ... covers covers. stegos. . . . .. covers .. . . . V.1 . ... stegos. . . . .. stegos .. . . . V.1 . ... Model training. Covers+pixels off. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 12/22.
(18) Introduction and background. Pixels-off technique. Conclusions. Phase2 : stegos generation Phase2: stegos generation Phase3 : training with data augmanation covers. . . . . . .. ..covers .. . . ...V.1 . ... Data embedding. stegos. . . . . . .. stegos .. .. . . ...V.1 . ... covers covers. steg. . . . .. covers .. . . . V.1 . ... stegos. Model training. Protocol I Chose an embedding algorithm, I Set a payload size, I Two sets of stegos are generated. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 13/22.
(19) Introduction and background. Pixels-off technique. Conclusions. Phase2 : stegos generation Phase2: stegos generation. The embedding modification probabilities map for the cover (resp."pixelsoff" version) used by S-UNIWARD model with a payload of 0.4 bpp Phase3 : training with data augmanation. covers. . . . . . .. ..covers .. . . ...V.1 . ... Data embedding. stegos. . . . . . .. stegos .. .. . . ...V.1 . ... covers covers. steg. . . . .. covers .. . . . V.1 . ... stegos. Model training. Protocol I Chose an embedding algorithm, I Set a payload size, I Two sets of stegos are generated. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 13/22.
(20) Introduction and background. Pixels-off technique. Conclusions. Phase2 : stegos generation Phase2: stegos generation. The embedding modification probabilities map for the cover (resp."pixelsoff" version) used by S-UNIWARD model with a payload of 0.4 bpp Phase3 : training with data augmanation. covers. . . . . . .. ..covers .. . . ...V.1 . ... Data embedding. stegos. . . . . . .. stegos .. .. . . ...V.1 . ... covers covers. steg. . . . .. covers .. . . . V.1 . ... stegos. Model training. Visualization by elevation for the PMs differences. Protocol I Chose an embedding algorithm, I Set a payload size, I Two sets of stegos are generated. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 13/22.
(21) Introduction and background. Pixels-off technique. Conclusions. Global flowchart of the pixel-off technique. Phase1: covers preparation. Phase2: stegos generation Phase3: training with data augmentation. covers. Pixels-off. covers. covers. +. . . . . . .. ..covers .. . . ...V.1 . ... covers. . . . . . .. ..covers .. . . ...V.1 . ... Data embedding. stegos. . . . . . .. ..stegos .. . . ...V.1 . ... covers covers. stegos. . . . .. covers .. . . . V.1 . ... stegos. . . . .. stegos .. . . . V.1 . ... Model training. Covers+pixels off. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 14/22.
(22) Introduction and background. Pixels-off technique. Conclusions. Phase3 : training with pixels-off enrichment. Phase3: training with data augmentation. Protocol. covers. covers. stegos. . . . .. ..covers .. . V.1 . ... stegos. VA. model training. Mehdi Yedroudj. . . . .. stegos .. . . . V.1 . ... I Prepare the training set (initial cover/stego database + the pixels-off cover/stego images), I Choose whether to use other given data augmentation techniques e.g. VA, I Initiate the model training.. Pixels-off: Data-augmentation Complementary Solution. 15/22.
(23) Introduction and background. Pixels-off technique. Conclusions. Experimental results Setup 1: The enrichment method: pixels-off, The steganalysis: Yedroudj-Net, The database: BOSSbase.. B = BOSS B1 = B+100-off B2 = B1 +256-off B3 = B2 +400-off B4 = B3 +1024-off. WOW 0.2bpp 0.4bpp 27.71 15.27 25.31 14.3 23.95 13.41 23.5 13.44 23.8 13.65. SUNIWARD 0.2bpp 0.4bpp 35.42 22.70 33.1 19.4 29.8 17.8 29.3 16.95 29.2 16.98. examples (pairs) 4,000 8,000 12,000 16,000 20,000. conv time 4-5h 9-10h 13-14h 17-18 21-22. Conclusion: I The optimal parameter is roughly around P = 400 pixels-off, I Combining various enrichments with P between 100 and 1024 improves the steganalysis efficiency. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 16/22.
(24) Introduction and background. Pixels-off technique. Conclusions. Experimental results Setup 2: The enrichment method: pixels-off, The steganalysis: CovPool-Net, The database: BOSSbase.. B = BOSS B1 = B+100-off B2 = B1 +256-off B3 = B2 +400-off B4 = B1 VA. WOW 0.2bpp 0.4bpp 26.08 15.60 25.33 14.63 24.88 13.11 23.34 13.02 17.5 9.23. SUNIWARD 0.2bpp 0.4bpp 31.89 18.32 28.54 16.25 26.61 15.00 26.64 15.44 21.58 10.54. examples (pairs) 4,000 8,000 12,000 16,000 64,000. conv time 5-6h 10-11h 14-15h 19-20h 10-11h. Conclusion: I The proposed method can improve performance of different steganalyzer, I Accumulating VA + pixels-off can improves further the performance. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 17/22.
(25) Introduction and background. Pixels-off technique. Conclusions. Does other weak noise signal work? Setup 3: The enrichment method: pixels-off, Gaussian, salt&pepper noise, The steganalysis: Yedroudj-Net, The database: BOSSbase. BOSS BOSS+100-off BOSS+Gaussian BOSS+salt&pepper (d = 0.05) BOSS+salt&pepper (d = 0.0016). WOW0.4 15.27 14.3 16.08 15.16 14.76. SUNIWARD0.4 22.70 19.4 23.25 22.25 19.92. Conclusion: I Low-power noise (less than 1.5% modified pixels) can be useful, I Other noises such as +/-1 noise achieve good results. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 18/22.
(26) Introduction and background. Pixels-off technique. Conclusions. SCA-database: Setup 4: The enrichment method: selective pixels-off, The steganalysis: Yedroudj-Net, The database: BOSSbase. BOSS BOSS+100 off BOSS+100 off-lowP BOSS+100 off-highP. WOW0.4 15.27 14.3 15.17 13.65. SUNIWARD0.4 22.70 19.4 20.85 18.15. Conclusion: I More beneficial to limit the pixels-off to pixels with a high modification probability, I Another way of doing SCA steganalysis, by generating SCA training sets (to be investigated). Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 19/22.
(27) Introduction and background. Pixels-off technique. Conclusions. Outline. Introduction and background Pixels-off technique Conclusion. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 20/22.
(28) Introduction and background. Pixels-off technique. Conclusions. Conclusion The pixel-off technique is: I A novel technique for data-base enrichment for CNN-based steganalysis. I Close in principle to noise addition, but made so that the pixel distribution of the resulting image remains close to that of the original image. I Efficient, simple to implement, and come with low complexity. I Suitable to be a complementary option to other enrichment techniques. I May be used for building informed database ”Side-Channel-Aware database”.. Mehdi Yedroudj. Pixels-off: Data-augmentation Complementary Solution. 21/22.
(29) Pixels-off: Data-augmentation Complementary Solution for Deep-learning Steganalysis Thank you for your attention IH&MMSec2020 15 June 2020.
(30)
Documents relatifs
The results are even better when we apply both augmentation techniques (database enhancement and virtual augmentation); the error probability is reduced by 5-7% compared to
Our JDBC proxy driver provides a minimally-intrusive way to augment data sci- ence workflows with additional functionality. The currently implemented driver plugins support
Smote creates pseudo instances based on the neighborhood which means pseudo instances of women who belong in negative class will have higher probability to include properties of
On IMDb, the BANKS-I versions are the top performing algorithms, followed by BANKS-II, DPBF, and DISCOVER-II; in this case the UNIPD variants perform very simi- larly to the
Data augmentation from RGB to chlorophyll fluorescence imaging Application to leaf segmentation of Arabidopsis thaliana from top view images.. Conference on Computer Vision and
Nous exposons dans le présent travail neuf cas de luxation de l’épaule type Erecta, colligés au service de traumatologie orthopédie au CHU Ibn Sina de Rabat sur
While the Jaccard index of the BG class is equally improved using different data augmentation techniques, the networks trained with color transfer give better results for the
We immediately notice that at both adaptation speeds our models are capable of surpassing the all-data model 11 in terms of accuracy on the new words, despite having seen fewer than
