1 / 25

Breaking HUGO the Process discovery Jessica Fridrich , Jan Kodovský , Vojtěch Holub , and Miroslav Goljan

Breaking HUGO the Process discovery Jessica Fridrich , Jan Kodovský , Vojtěch Holub , and Miroslav Goljan. Presented by: Landry EDI, Na Yi. Introduction. 13th Information Hiding Conference , Prague, Czech Republic, May 18–20, 2011

hall
Download Presentation

Breaking HUGO the Process discovery Jessica Fridrich , Jan Kodovský , Vojtěch Holub , and Miroslav Goljan

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Breaking HUGOthe Process discoveryJessica Fridrich, Jan Kodovský, VojtěchHolub, and MiroslavGoljan Presented by: Landry EDI, Na Yi

  2. Introduction • 13th Information Hiding Conference, Prague, Czech Republic, May 18–20, 2011 • HUGO (Highly Undetectable steGO): one of the most advanced steganographic systems ever published. • Description of the author’s experience with the BOSS competition in chronological order

  3. Purpose of the competition • Purpose: Identify stego images from BOSSbase images That embedded 0.4bpp payload using HUGO • Performance evaluation based on a single scalar value called the BOSSrank score which is the percentage of detection accuracy defined as: where PFA and PMD are the probabilities of false alarm and missed detection. When the score is computed from BOSSrank images, it will always be referred to as the “BOSSrank score.” • Obtain more than 80% of BOSSrank score

  4. Early Ideas-the public selection channel Is the public selection channel a problem? • Selection Channel: placement of embedding changes • The first idea: To utilize the fact that the attacker can approximately determine the probabilities with which each pixel was changed during embedding. • A weakness of adaptive embedding that may be exploited is to reveal where embedding changes are made and where they are not. • HUGO, which source code is public, modifies pixel xiby ±1 with probability pXithat can be determined from the cover image X and the payload.

  5. Figure 1: pYivs.pXi Early Ideas-the public selection channel • However, the probabilities computed from the stego image will be slightly different pYi≠ pXi. • Fig.1 shows probability of embedding change computed from the stego image, pYivs.pXi(for image no. 50 from BOSSbase). • pXi≈ pYi with the largest relative errors for small pXi. • In particular, |pYi− pXi| ≤ 0.05 for 99.4% of pixels, |pYi− pXi| ≤ 0.01 for 85.2% of pixels, and |pYi− pXi| ≤ 0.001 for 40.4% of pixels.

  6. Early Ideas-the public selection channel • As the payload is known and pXi≈ pYi , in theory one could derive the values of cover-image statistics, such as histograms or co-occurrence matrices. But using these estimate statistics for steganalysisstill be problematic. • Result(why this idea didn’t work?): Because HUGO does not introduce any easily detectable changes and one may not have any way of telling whether they are inspecting a cover or a stego image.

  7. Early ideas - detection by correlation • Motivation: If one were able to estimate from the stego image whether a given pixel was modified by 1 or −1 with probability, one could detect HUGO (and ±1) embedding using a correlation. • Result(why this idea didn’t work?): As the correlation varies greatly across images, it’s impossible to find a threshold for it . For example, in some images, the increase in correlation up to 60% but the average increase was only 1.74%, which is by several orders of magnitude smaller than the variations of the correlation value across images.

  8. Is pixel domain useful? • Motivation: Features computed from differences between neighboring pixels will lead to weak detection, because the embedding algorithm was designed to preserve statistics in a 107-dimensional feature space built from joint statistics of pixel differences on 7 × 7 neighborhoods. • Direction: To compute features in an alternative domain, such as the wavelet domain. To this end, the team decided to modify the WAM(Wave Model) feature vector.

  9. Is pixel domain useful? • Improvement of WAM: • Replacing the Wiener filter with its adaptive version in which the fixed noise variance was replaced with the variance σ2w of the stego noise si at pixel i, σ2w,i = pYi . • The best performance was achieved by merging the original 27 WAM features, 27 content features, and 27 WAM features obtained using the adaptive filter and adding to it the same set of 81 features from a re-embedded image (total of 162 features WAMCPre).

  10. Is pixel domain useful? • The final performance is summarized in Table 1 shown as below: Table 1: performance of the WAM steganalyzer and its various extensions.

  11. Going back to pixel domain • Motivation: The best detection is usually achieved by forming features directly in the embedding domain. Since this is where the embedding changes are localized and thus most pronounced. The key idea and a major breakthrough in our effort to break HUGO was the realization that another way to form a statistic that spans more than four pixels is to use higher-order pixel differences (residuals).

  12. 80% milestone • The 80% milestone was within reach by the end of October. • To be able to train a classifier at this scale, the teamhad to abandon SVMs and reinvent the entire machine learning approach. But before they get to that, in the next section we will describe the Warden’s nightmare.

  13. Dreaded cover-source mismatch • Steps: find the optimal value of the threshold T, add other versions of the features and train on a larger number of images. • moved to a four–dimensional co-occurrence operator for the QUANT feature set, obtaining thus a 4,802-dimensional feature vector (2×(2×3+1)4=4802) • As result while they steadily improved detection accuracy on BOSSbase by adding more features, the BOSSrankscore was moving in the opposite direction: facing the dreaded cover-source mismatch issue • The detector lacked what is recognized in detection theory as robustness

  14. Dreaded cover-source mismatch • Another way to increase robustness was to train on a larger set of images • Another set of 6,500 images taken in raw format by 22 different cameras and converted using the same script that was used for creating the BOSSbase Were Added to BOSSbase 0.91: seemed to make the score worse

  15. Dreaded cover-source mismatch • The cover-source mismatch has been recognized by the research community as a serious issue that may complicate deployment of steganalysis in rea l life because the performance of the WAM steganalyzer on images could be vastly improved if the steganalyzer was trained on images from the exact same camera or, to a slightly lesser degree, on images from a camera of the same model • Decision of finding out as much as possible about the source of covers for BOSSrank to improve the BOSSrank score.

  16. Forensic analysis of BOSSrank • On October 14, extraction of the sensor fingerprint for each camera from BOSSbase, then, test of all BOSSrank images for the presence of the fingerprints • The LEICA M9 fingerprint was found in approximately 490 images • Based on visual inspection, some portion of images was taken in the Pacific Northwest (Cars license plates)

  17. BOSSrank source identification Panasonic Lumix DMC-FZ50 had been used, and it belonged to Tomáš Filler

  18. New Orientation • More images from LEICA M9 Camera • Rented for a week (October 23–30) because too expensive ($7000) • 7,301 images in their original resolution of 18 Mpx was taken and processed using the BOSS conversion script and subsequently embedded with payload 0.4 bpp • After extrating the MINMAX+QUANT features, they built two detectors: one G-SVM trained on all BOSSbase images that would be used for detection of all non-Leica images from BOSSrank, and the second G-SVM specifically trained on the union of their 7,301 Leica images and the 2,267 Leica images from BOSSbase • As result, 74% as BOSSrank score and no improvement after running a couple of more experiments • The content of images has obviously a major influence on content- adaptive steganalysis. The cover source is a very complex entity that is affected by the lens, the environment in which pictures are taken and even the photographer’s habits

  19. Diversity • Refocusing effort to creating a more diverse feature set while keeping the dimensionality around 3,000, what seemed as a sweet spot for the given training set. To this end, they used a lower threshold T = 3 and incorporated higher-order differences among neighboring pixels. One can easily extend the MINMAX and QUANT feature vectors to higher-order residuals: • The strategy was successful. By training a G-SVM on images from BOSSbase they obtained a BOSSranks core of 76% on November 13

  20. Way to breaking HUGO • instead of blindly increasing the threshold and co-occurrence order, they had to increase the feature diversity. • The complexity of training a G-SVM, however, was beginning to limit the speed of development, while the performance of the much faster L-SVMs was sub-par compared to G-SVMs . They needed an alternative machine learning tool that would enable faster development and testing of many ideas and combinations of features. • New Approach: an inexpensive , fast, and scalable machine learning approach

  21. The new Approach • Details, experiment and comparison with SVMs shown in 2nd paper • Start with a feature set of full dimensionality d, and build a simple classifier on a randomly selected subset of dred<<d features while using all training images • The classifier is a mapping: where 0 and 1 stand for cover and stego classes • Repeat L times with a different random subset of the features: L classifiers obtained given a feature vector b. • Decision based on fusion of the individual decisions of all L classifiers • the ensemble classifier can be trained on 2×9,074 images with a 10,000-dimensional feature set with L=31 and dred =1600 • decisions about the entire BOSSrank made in about 7min achieved on a DELL Precision T1500 machine with 8GB of RAM and 8 Intel Cores i7 running at 2.93GHz

  22. The final attack • The scalability and low-complexity of the ensemble classifier allowed improvement of the BOSSrank scores imply by gradually scaling up the features and training sets • On November 15, 77% was reached with a set consisting of 5,330 features trained with L=31 and dred =1600 on the entire BOSSbase • On November 29, they added more features to the 5,330-dimensional set to form a feature vector with 9,288 elements. • Added features: 1 ) the QUANT feature vector with q=2 constructed from fourth-order residuals and a 4D co-occurrence (dimensionality 2×625) formed from horizontal and vertical samples 2 ) an equivalent of the QUANT feature with q=2 constructed from second-order residuals and a 4D co-occurrence of residuals arranged into a 2×2 square (2×625) 3 ) a vector constructed from residuals computed using a translationally-invariant Ker–Böhme kernel as followed: And a 3D co-occurre nce (729 features) and the same co-occurrence after quantizing the residual with q=2 (another 729 features ) • All together, 5330+2500+1458=9288 features. When tr ained on BOSSbase, this set produced a score of 76%

  23. Scores The rest of the record submissions are displayed bellow. The last three were achieved with L=51,51,71 and dred =2400

  24. Conclusion • First, there is no reason why steganalysts should frown at high-dimensional feature sets. To the contrary, we believe that high-dimensional features are a necessity to attack advanced steganography. The dimensionality could probably be reduced by clever marginalization, however, automatized design using ensemble classifiers is preferable to hand-crafting the features. The ensemble classifiers offer a scalable and quite simple classification with very similar performance to that of the much more complex SVMs

  25. Thank you ! Questions?

More Related