YASS Yet Another Steganographic Scheme that Resists Blind Steganalysis K. Solanki*, A. Sarkar +, and B. Manjunath Vision Research Laboratory Department of Electrical and Computer Engineering University of California, Santa Barbara, CA 93106, USA * * Mayachitra, Inc Hollister Ave. Santa Barbara, CA-93111, USA
Steganography: Art and science of covert communication Steganographic security: Cachin’s criteria Kullback-Leibler divergence smaller than epsilon Inherently assumes availability of “natural” distributions Statistical steganalysis Suspected stego signal evaluated against assumed or computed cover distribution June 11,
Steganalysis: Winning the battle? Blind Statistical steganalysis Uses supervised learning on specific image features Self-calibration mechanism used to ensure that features capture changes due to embedding only. E.g. Cropping a few pixel rows and/or columns Recent results close to perfect [Pevny and Fridrich ‘07]: can reliably determine which stego scheme was used (out of 5) with >95% accuracy June 11,
YASS: “Stirmark of Steganography” Idea: Desynchronize the steganalyst by randomizing the embedding locations Disables the self-calibration process But, must advertize or ship the image in a standard format (such as JPEG) Causes errors in the recovered bits Use erasures and errors correcting codes Previously employed for high-volume hiding June 11, A recipe for resisting blind steganalysis
A peek at the Results Can resist several recent steganalysis techniques JPEG steganalysis with self-calibration Pevny and Fridrich’s 23-dim features (SPIE’06) and their more recent 276-dim feature (SPIE’07) Farid’s 72-dim features DCT histogram-based features Spatial domain steganalysis Xuan et al’s 39-dim features based on wavelet characteristic functions (IH ‘05) JPEG steganalysis based on above features Chen et al’s 324-dim features (ICIP ‘06) June 11,
Outline Introduction Related Work Resisting Blind Steganalysis YASS for JPEG Steganography Results Discussion Future Work June 11,
Outline Introduction Related Work Steganography Steganalysis Resisting Blind Steganalysis YASS for JPEG Steganography Results Discussion Future Work June 11,
Related work JPEG Steganography: Schemes that match or restore marginal statistics Sallee’s model based methods (MB1 and MB2) Fridrich et al’s Perturbed Quantization (PQ) Solanki et al’s Statistical Restoration schemes OutGuess, StegHide Matrix embedding schemes (such as F5) Steganalysis schemes Many related schemes already listed earlier June 11,
Outline Introduction Related Work Resisting Blind Steganalysis How blind steganalysis schemes work Can we defeat them? How? YASS for JPEG Steganography Results Discussion Future Work June 11,
Blind Steganalysis: Key Ingredients 1.Self-calibration mechanism Used to estimate the cover image statistics from the stego image For JPEG steganography: Crop a few pixel rows or columns and recompress 2.Features capturing cover memory Most stego scheme hide data on a per-symbol basis Higher order dependencies harder to match or restore 3.Powerful machine learning Ensures that even the slightest statistical variation in the features is learned by the machine June 11,
Blind steganalysis is quite successful Self-calibration process is perhaps the most important ingredient Derived features are insensitive to image content, but quite sensitive to embedding changes Steganalysis successful in spite of unavailability of universal image models Results presented in [Pevny and Fridrich, SPIE ‘07] are close to perfect! June 11,
So, What can the steganographer do? Preserve all the features of the image… Is this practically feasible? How does it affect the embedding rate? But, lets not forget: The steganalyst must depend on the stego image to estimate the cover image statistics Way out: Embed data in a way that distorts the steganalyst’s estimate of the cover image statistics June 11,
Distorting Steganalyst’s Estimate Hiding with high embedding strength Cover image statistics can no longer be reliably derived from the available stego image Also observed and reported in recent work by Kharrazi, Sencar, and Memon (ICIP ‘06) Randomized hiding The algorithm to estimate the cover statistics can be effectively disabled Can randomize the hiding location, the choice of transform domain, the coefficient, or even the hiding method June 11, Disadvantages 1.Likelihood of high perceptual distortion 2.Possibility of data being detected by universal image models Our Choice! Simple implementation explored in this work: Hide in random locations
Outline Introduction Related Work Resisting Blind Steganalysis YASS for JPEG Steganography Results Discussion Future Work June 11,
YASS for JPEG Steganography Idea: Embed data in randomized block locations The blocks do not coincide with the JPEG 8x8 grid Errors caused due to initial JPEG compression after embedding data Use erasures and errors correcting codes June 11,
YASS Embedding: JPEG Grid June 11, pixels
YASS Embedding June 11, B x B block Here, B=10 B is called “big block size” Example grid used in embedding Randomized block location
Reduction in Embedding Rate Wasted real estate of the image Due to choice of bigger blocks Can reduce this wastage by putting more than one blocks in larger bigger blocks Eg 16 blocks in 34x34 sized block Errors due to initial JPEG compression Use erasures and errors correcting codes Previously employed in [Solanki et al, Trans. Image Pro. Dec 2004] June 11,
Advantages of Coding Framework Robustness against initial JPEG compression Enabling active steganography by providing robustness against distortion constrained attacks Allows choice of embedding locations to reduce perceptual distortion Do not hide in zero valued DCT coefficients June 11,
Outline Introduction Related Work Resisting Blind Steganalysis YASS for JPEG Steganography Results Discussion Future Work June 11,
Results Experimental Set-up Embedding volume Performance against steganalysis Comparison with competing methods Control experiment: Comparison with standard hiding at same rate June 11,
Experimental Set-up June 11, Database of > 2000 images Hide Data Training Testing Use Supervised learning for steganalysis Two Datasets: TIFF images JPEG images
Results Experimental Set-up Embedding volume Performance against steganalysis Comparison with competing methods Control experiment: Comparison with standard hiding at same rate June 11,
Embedding volume June 11, Embedding volume for some standard 512x512 images: QF h = design quality factor used for embedding QF a = output JPEG quality factor Number of bits are the number of information bits providing error-free recovery
Hiding Rate for Lenna Image June 11, bpnc = bits per non-zero coefficients
Results Experimental Set-up Embedding volume Performance against steganalysis Comparison with competing methods Control experiment: Comparison with standard hiding at same rate June 11,
Steganalysis schemes Farid: 72-dimensional feature vector based on moments in the wavelet domain PF-23: Pevny and Fridrich’s 23-dimensional DCT feature vector PF-274: Pevny and Fridrich’s 274-dimensional feature vector that merges Markov and DCT features DCT hist.: Histogram of DCT coefficients from a low- frequency band Xuan-39: Spatial domain steganalysis proposed by Xuan et al Chen-324: JPEG steganalysis based on statistical moments of wavelet characteristic functions June 11,
JPEG dataset June 11,
June 11, TIFF dataset: Similar results
Results Experimental Set-up Embedding volume Performance against steganalysis Comparison with competing methods Control experiment: Comparison with standard hiding at same rate June 11,
Comparison with competing methods Compared the method with OutGuess and StegHide Ensured that the hiding rates are equivalent Not straightforward since YASS uses error correcting codes The code rate 1/q determines the embedding rate Tested for q=10 and q=40 June 11,
Comparison with OutGuess and StegHide June 11, YASS is undetectable for most configurations for these steganalysis schemes, while competing methods are detectable at the same hiding rates! F5, using matrix embedding, performs equivalent to YASS at equivalent hiding rates
Results Experimental Set-up Embedding volume Performance against steganalysis Comparison with competing methods Control experiment: Comparison with standard hiding at same rate June 11,
Control experiment: Standard hiding Is the good performance of YASS simply due to lowered embedding rate? Control experiment: Compare with naïve hiding scheme Hide in random frequency locations (termed RF scheme) Trivial embedding efficiency of 2 bits/coefficient Tested hiding rates 2 bits per 64 (8x8 block), 1 bit per 64, and 1 per 128 coefficients June 11,
YASS verses RF June 11, It can be seen that the naive RF scheme performs quite well, however, the performance of YASS is consistently better.
Outline Introduction Related Work Resisting Blind Steganalysis YASS for JPEG Steganography Results Discussion Future Work June 11,
Change Rate and Embedding Efficiency Change rate: Number of coefficients modified during the embedding process Embedding efficiency: Number of bits hidden per coefficient changed Change rate is encoder’s budget; it can be “used” To provide robustness by using redundancy and giving up some embedding efficiency To improve the embedding efficiency via matrix embedding but causing an increase in the fragility of the system YASS goes the first way… June 11,
YASS and Matrix Embedding The hiding rate of YASS is relatively low At these hiding rates, equivalent level of steganographic security can be achieved by using matrix embedding An advantage of YASS is that it can provide robustness against distortion constrained attacks June 11,
Outline Introduction Related Work Resisting Blind Steganalysis YASS for JPEG Steganography Results Discussion Future Work June 11,
Future Work Improve the embedding rate Test for an active steganographic framework Explore other avenues of randomization Embedding method Transform domain How can this scheme be detected? June 11,
Thank you MATLAB code available at