Journal of VLSI Signal Processing 41, 293303, 2005 c 2005 Springer Science + Business Media, Inc.

Manufactured in The Netherlands. DOI: 10.1007/s11265-005-4153-1

Data Masking: A New Approach for Steganography?

REGUNATHAN RADHAKRISHNAN, MEHDI KHARRAZI AND NASIR MEMON Polytechnic University, Brooklyn, NY 11201, USA

Abstract. It is well known that encryption provides secure channels for communicating entities. However, due to lack of covertness on these channels, an eavesdropper can identify encrypted streams through statistical tests and capture them for further cryptanalysis. Hence, the communicating entities can use steganography to achieve covertness. In this paper we propose a new form of multimedia steganography called data masking. Instead of embedding a secret message into a multimedia object, as in traditional multimedia steganography, we process the entire secret message to make it appear statistically similar to a multimedia object itself. Thereby we foil an eavesdropper who is primarily applying statistical tests to detect encrypted communication channels. We show that our approach can potentially give a covert channel capacity which is an order of magnitude higher than traditional steganography. Our experiments also show that steganalyzers trained with stego objects of known steganographic have a low detection accuracy for datamasked multimedia objects.

1.

Introduction

Rapid developments in cryptography, cryptographic standards, and relaxed export controls on encryption schemes have made cryptographic software widely available, user friendly and in some cases transparent to end-users. The relative ease with which secure channels of communication can be created between two parties using cryptography, coupled with the explosion in network trafc, has posed a problem to the surveillance operations typically carried out by law enforcement and intelligence agencies. In order to keep up with the growth in communication trafc and secure channels, massive eavesdropping campaigns, such as Carnivore and Echelon, have been developed. Carnivore, for example, is a combination of hardware and software, which collects all electronic packets that are sent to and from the Internet Service Provider where it is installed. Following this collection procedure, several lters are applied to isolate emails with certain keywords which could be further analyzed off-line. These lters could potentially include the performance

of statistical tests to isolate encrypted emails for offline cryptanalysis. Pervasive eavesdropping like that performed by Carnivore leads to the need for covert channels, in which communicating entities hide their communication channel itself from eavesdroppers. Steganography refers to the science of invisible communication. Unlike cryptography, where the goal is to secure communications from an eavesdropper, steganographic techniques strive to hide the very presence of the message itself from an observer [1]. As broadband technologies improve bandwidth at the last-mile, multimedia content, such as still images, audio and video, have gained increasing popularity on the Internet. Given the high degree of redundancy present in a digital representation of multimedia content (despite compression), there has been an increased interest in using multimedia content for the purpose of steganography. Indeed many such techniques have been devised in the past few years. For an excellent survey of such techniques, the reader is referred to [2]. Briey, the general model for steganography, illustrated in Fig. 1, where we have

294

Radhakrishnan, Kharrazi and Memon

Figure 1. Framework for Secret Key Passive Warden Steganography. Alice embeds secret message in cover image (left). Wendy the warden checks if Alices image is a stego-image (center). If she cannot determine it to be so, she passes it on to Bob who retrieves the hidden message based on secret key (right) he shares with Alice.

Alice wishing to send a secret message m to Bob. In order to do so, she embeds m into a cover-object c, to obtain the stego-object s. The stego-object s is then sent through the public channel. Wendy who examines all messages in the channel, should not be able to distinguish in any sense between cover-objects (objects not containing any secret message) and stego-objects (objects containing a secret message). In this context, steganalysis refers to the body of techniques that aid Wendy in distinguishing between cover-objects and stego-objects. It should be noted that Wendy has to make this distinction without any knowledge of the secret key which Alice and Bob may be sharing and sometimes even without any knowledge of the specic algorithm that they might be using for embedding the secret message. There are two problems with multimedia steganography that seem to have been ignored in the literature. Firstly, in multimedia steganography, the size of the covert message is relatively much smaller than the size of the multimedia object (the stego object) that carries the covert message. In fact recent developments in steganalysis techniques have demonstrated that embedding a secret message in a multimedia object, like say a still image, causes certain statistically discernible artifacts that reveal the presence of the secret message thereby exposing the covert communication channel [36]. Such developments further limit the number of bits that can safely be hidden in a multimedia object from sophisticated steganalysis techniques. This leads to a cover object that is one or more orders of magnitude larger than the secret message, reducing the effective bandwidth of covert communication channels.

The second problem is that the common interpretation of the warden-based framework when used for multimedia steganography described has been that the warden examines the cover-object (or stego-object) perceptually and statistically. However, perceptual tests, which may have been practical on low volume communication channels, would not be practical with high volume communication channels. Such channels must rely on automated statistical tests for detecting potential stego objects and only then may apply (if at all) some form of perceptual tests to the selected candidates. Indeed, large scale eavesdropping operations like Carnivore rely rst on statistical analysis to isolate interesting messages. Hence, if Alice and Bob can evade statistical tests performed by Wendy, they may not have to undergo any perceptual test at all! So for example, if the stego object is statistically indistinguishable from an image, its perceptual quality does not matter at all. Can this fact be exploited to improve covert communications in any sense? In this paper we show that this is indeed possible. Specically, we propose a novel technique, which we call Data Masking, for multimedia steganography. We again look at the problem of Alice and Bob wishing to communicate covertly, but we assume that the warden Wendy is an entity like Carnivore and will rst examine messages statistically. Only if the statistical tests indicate a possibility of covert communication, would she then examine the message further. Hence instead of selecting a cover object and embedding a covert message in it like traditional steganography, we mask the secret message to make it appear like a normal multimedia object under typical statistical examinations.

Data Masking: A New Approach for Steganography?

295

Figure 2.

Audio datamasking using spectral factorization.

To demonstrate our approach we assume the covert message is encrypted and hence a random bit stream. Since a random bit string can be easily identied by Wendy using simple statistical tests, we process the random bit stream to make it appear like a audio signal or image, in a statistical sense. This process removes randomness from the encrypted message, and makes it difcult for Wendy to distinguish it from a typical audio stream or image. We refer to the process of converting a random bitstream into audio and image as audio datamasking and image datamasking respectively. The rest of the paper is organized as follows: in the following section we describe the proposed system for audio and image datamasking. Section 3 presents some experimental results and possible improvements and we conclude in Section 4 with future work. 2. Proposed System

Audio datamasking using spectral factorization. Audio datamasking using LPC analysis-synthesis. 2.1.1. Audio Datamasking Using Spectral Factorization. Audio datamasking using spectral factorization is illustrated in Fig. 2. Let us consider the cipher stream as samples from a Wide Sense Stationary (WSS) Process, E. We would like to transform this input process with high degree of randomness to another stationary process, A, with more correlation between samples by using a linear lter, H. It is well known that Power spectrum of a input WSS Process, A(w ), to a linear time invariant system and Power spectrum of the corresponding output Process, E(w ) are related by the following equation: E ( w ) = | H ( w ) |2 A ( w ) (1)

If Alice wants to send an encrypted message to Bob, the warden Wendy would be able to detect such a message as an encrypted stream since it would exhibit properties of randomness. In order for a secure channel to achieve covertness, it is necessary to preprocess the encrypted stream at the end points to remove randomness such that the resulting stream defeats statistical tests for randomness and the stream is reversible at the other end. In this section, we propose schemes that make the processed encrypted streams appear like more correlated multimedia streams (audio and image). 2.1. Audio Datamasking

If E(w ) is a white noise process, then H(w ) is the whitening lter or Wiener lter. Since the encrypted stream is random, its power spectral density is at and resembles the power spectral density of a white noise process. Then, the desired Wiener lter can be obtained by spectral factorization of (E(w )/A(w )) followed by selection of poles and zeros to obtain the minimum phase solution for H(w ). Since the factor (E(w )/A(w )) can have arbitrary shape, it would require a lter of very high order for realization. 2.1.2. Audio Datamasking Using LPC AnalysisSynthesis. Audio datamasking can also be performed using LPC Analysis/Synthesis as shown in Fig. 3. The LPC Analysis lter for reference Audio clip, A, is obtained as follows. Let X 0 , X 1 , X 2 . . . X N 1 represent N previous samples of the reference audio clip. The goal is to obtain the lter coefcients h 0 , h 1 , h 2 . . . h N 1

We propose the following two methods for audio datamasking.

296

Radhakrishnan, Kharrazi and Memon

Figure 3.

Audio datamasking using LPC analysis/synthesis.

such that E (( X i X i )2 ) is minimized. Here X i is the predicted value of the current sample based on N previous samples in the reference audio and is dened N 1) as ( k =0 h k X k . Using the orthogonality principle (Hilbert space projection theorem), N equations (called Yule-Walker equations) can be set up to solve for the optimal lter coefcients in the MMSE (Minimum Mean Square Error) sense. Then, the inverse of the LPC analysis lter so designed, can be used to lter the noise-like cipher stream to remove randomness from cipher stream and transform it into a reference audio-like waveform that has more correlation between samples. With the knowledge of lter coefcients the receiver can reconstruct the cipher stream from the reference audio, as in the Inverse Wiener ltered cipher stream.

2.2.

Image Datamasking

Since the source model for speech is well understood, LPC Analysis-synthesis could be used for datamasking. The all-pole synthesis lter is known to model the human vocal tract lter. However, the source model for images is not as well understood and hence we propose a different scheme for image datamasking as shown in Fig. 3. As in audio datamasking, we use the prediction error to mask the encrypted stream. Each pixels intensity value is predicted based on its causal neighborhood using a MED predictor that is employed in the JPEG-LS lossless compression standard [8]. The causal neighborhood consists of W, NW, N, NE (west, northwest, north, northeast) pixels. According to the MED predictor, the predicted value of current pixel is

min (W, N) if N W max(W , N ) or it is max(W , N ) if N W min(W , N ) or it is W + N NW otherwise. A Huffman codebook, that is constructed for the Laplacian prediction error probability density function (PDF), is used to decode the input random encrypted stream. Then the PDF of Huffman decoded encrypted samples would also have a Laplacian PDF. Hence, one can modify each pixel so that the prediction error at that pixel is matched with the Huffman decoded encrypted stream sample. Note that in the proposed system for image datamasking, there are two parameters that the communicating parties should exchange beforehand. The rst is a Huffman codebook designed for coding prediction errors with a Laplacian PDF and the second is a threshold to select the range prediction error values to be mapped to encrypted stream samples. For instance, if the chosen threshold in M, there would 2M+1 huffman codewords mapping { M . . . 0 . . . M }. This threshold controls the tradeoff between datamasking capacity and quality of the data masked image. By choosing the threshold small, one would select only smooth regions of the image to hide information and hence the quality of the datamasked image would be better. With the knowledge of these two parameters, the receiver would map the prediction error at each pixel to the random encrypted stream. In the following section, we present some of our experimental results to evaluate the performance of audio and image data masking. 3. Experimental Results

We chose an AES (Advanced Encryption Standard) Advanced Encryption Standard [7], encrypted stream

Data Masking: A New Approach for Steganography?

297

Figure 4.

Proposed system for image datamasking.

Figure 5. Comparison of reference audio clip and encrypted stream in time domain.

Figure 6. Comparison of power spectral densities (PSDs) of reference audio clip and encrypted stream.

as input to our system. The systems desired output is to generate a multimedia object from this stream. We use any audio clip or image as a reference to guide the system in the process. 3.1. Results on Audio Datamasking

We picked a speech clip of duration 2 seconds as our reference audio, A. We read every seven bits from the encrypted stream and append a random LSB and interpret it as a sample from a random process, E. Figures 5 and 6 compare the encrypted stream and reference audio clip in time domain and frequency domain respectively. The reference audio is then segmented into frames of length 1024 samples, such that the time duration is short enough for the stationarity assumption to hold. Then for each frame of audio, an equal number of bytes

from the encrypted stream are read. A LPC Analysis lter of order 31 was designed, the inverse of which was used to lter the encrypted stream. We thus have a time-varying inverse Wiener lter shaping every 1024 samples of the encrypted stream to match the reference audio frame. Figure 7 shows the results of inverse Wiener ltering on encrypted stream corresponding to 2 s of reference audio clip and its PSD. It is clear that the ltered signal has more correlation than the encrypted signal and would potentially defeat most statistical tests for randomness. However, the ltered signal does not match the reference audio waveform exactly and hence the resulting waveform was noisier when listened to (perceptual test). If the shaping of the encrypted stream perfectly matched that of reference audio for each frame, then the inverse Wiener ltered

298

Radhakrishnan, Kharrazi and Memon

Figure 7.

Inverse wiener ltered encrypted stream & its PSD.

Figure 8. Reconstructed encrypted stream & its difference from original encrypted stream.

encrypted stream would also sound like the reference audio. In order to reconstruct the encrypted stream at the receiver, knowledge of LPC analysis lter coefcients for the current frame is essential. Therefore, the lter coefcients for the current frame are appended to the encrypted stream that was shaped in the previous frame and the coefcients corresponding to the rst frame were included in the header of the .wav le. Wiener ltering can then be performed for each frame by retrieving the ltering coefcients reconstructed from the previous frame to obtain the encrypted stream for the current frame. Figure 8 shows the reconstructed encrypted stream and its difference from the original. The maximum difference between the samples of original and reconstructed encrypted stream is within 0.015. Since each sample was interpreted as a byte from the encrypted stream and scaled by (1/64) and shifted by 1, a bit ip in the encrypted stream can occur only if the error is greater than (1/64). Therefore, if the error is within (1/64) or (0.0156), it is possible to reconstruct the encrypted stream without any bit error which is critical for proper decryption. Now that we are convinced that the proposed scheme would mask encrypted data from statistical tests, let us evaluate the performance of this scheme as a audio data hiding technique. If we assume that each lter coefcient occupies a word, we have an overhead of 124 bytes (4 31) for every 1024 bytes of encrypted stream. The effective payload, (1024124128) = 772 bytes, can be thought of as embedding capacity for this scheme per frame of audio. It has been observed that the quantization error introduced by a precision

of 8 bits per sample of the ltered encrypted stream would result in an error greater than (1/64). Therefore, we use 16 bit precision for each sample in the ltered encrypted stream and assume that the transmitted inverse Wiener ltered stream was received without any distortion. We can compare our schemes embedding capacity to that of LSB embedding, in which one bit is embedded in the LSB per sample of audio. The proposed scheme hides 6176 bits (8 772) per 1024 bytes of audio data whereas LSB embedding (which also has low robustness) can hide 1024 bits per 1024 bytes of audio data. Therefore, the embedding capacity of proposed scheme is 6.0313 times that of the capacity of LSB embedding. In order to achieve an embedding capacity of 6176 bits per 1024 bytes using LSB embedding technique, it would require changing at least 6 bits of the 8 bits in each sample! Note that in practice, steganalysis techniques given in [46] can detect the presence of LSB embedding even when 10% of the bits in the cover message have been ipped (in some cases this can be as low as 2%;). So we potentially have an order of magnitude larger capacity as compared to traditional multimedia steganography. Of course, we gain this by assuming that warden does not examine the message perceptually, but only statistically. In order to see what kind of statistical tests can detect the datamasked audio signals, we tested them with a number of existing audio steganalysis tools [9]. Table 1 summarizes the performance of some steganalysis techniques on 66 datamasked audio streams of total duration 29 min. It can be observed that all of the detectors trained with stego objects of known steganographic techniques have a low detection

Data Masking: A New Approach for Steganography?

299

Table 1. Detection performance of classiers trained with stego audio les of specic embedding techniques with embedding capacity of 6.03bits per audio sample; [A]: Number of actual datamasked audio streams; [B]: Number of audio streams detected by the ste[C ]) ganalyzer; [C]: Number of false alarms; Precision: ([ B ] ; Recall: [B] ([ B ][C ]) . [ A] [A] DSSS FHSSS Echo DctWHas Steganos S-tools 66 66 66 66 66 66 [B] 100 38 36 20 17 54 [C] 46 12 7 9 4 15 Precision (%) 54.0 68.42 80.56 55 76.47 72.22 Recall (%) 81.82 39.39 43.94 16.67 25.76 59.09

rate for the datamasked audio streams. This suggests that the artifacts introduced to the cover audio signal by the datamasking process has a different signature than that introduced by existing steganographic techniques. Hence, the classiers trained with stego audio les of specic embedding techniques were less successful with a dataset consisting datamasked audio les. 3.2. Results on Image Datamasking

A reference image was chosen and a Huffman codebook was designed for the prediction error PDF. The reference images gray scale values were re-mapped

to a smaller range so that the modied pixels value is still within [0 255]. Figure 9 shows the original reference image and the PDF of the prediction error. We used a threshold of 50 on the absolute value of prediction error and hence there were 101 huffman codewords to map the encrypted stream to prediction errors. Figure 10 shows the datamasked image and its prediction error PDF. The embedding capacity of this datamasking scheme for the chosen image was found to be 5.47 bits per byte of image stream. Figure 11 shows the tradeoff between embedding capacity and quality of the datamasked image for two different values of M. For all of the chosen M values, the embedding capacity is high compared to many of the multimedia steganographic schemes. With such a high embedding capacity, would the datamasked image escape statistical tests performed by existing steganalyzers? This motivated us to test a number of datamasked images with existing steganalyzers for known steganographic techniques. Tables 1 and 2 present the detection performance on two sets of datamasked images of a linear classier trained with stego images of specic embedding techniques. Dataset 1 contains 100 datamasked images with embedding capacity of 4.5 bpp & their corresponding 100 original images. Dataset 2 contains 71 datamasked images with the same embedding capacity & their corresponding original images. The original images in the Dataset 2 were chosen to be already noisy images.

Figure 9.

Original reference image and its prediction error histogram.

300

Radhakrishnan, Kharrazi and Memon

Figure 10.

Datamasked image and its prediction error histogram.

Figure 11.

Datamasked images with M = 25 and M = 15 with corresponding embedding capacities of 4.56 & 3.65 bpp.

The results from tables 1 and 2 show that most of the steganalyzers have a low detection rate for both datasets. Also, the false alarm rate is consistently higher for dataset 2 compared to dataset 1. Since the images in dataset 2 were already noisy before embedding, the detectors found it difcult to differentiate between the noise introduced by the embedding process and the noise already present, thereby increasing the false alarm rate. Since the choice of the cover image for steganography is upto the communicating parties,

the noisy images like those in dataset 2 would be more desirable for datamasking. In order to see the effect of embedding capacity on these two classiers, a similar experiment was performed on the same datasets but with a lower embedding capacity of 2.8 bpp. Tables 3 and 7 present the results for the testing dataset 1 with SVM classiers with two different embedding capacities. Tables 4 and 8 present the results of SVM classiers on dataset 2 with two different embedding capacities.

Data Masking: A New Approach for Steganography?

301

Table 2. Top Table: Detection performance of linear classiers trained with stego images of specic embedding techniques using Binary Similarity Metric (BSM) as features for Dataset 1 with embedding capacity of 4.5 bpp; [A]: Number of actual datamasked image streams; [B]: Number of image streams detected by the ste[C ]) ganalyzer; [C]: Number of false alarms; Precision: ([ B ] ; Recall: [B] ([ B ][C ]) . [ A] [A] F5r12 F5r11 Out Out+ Lsb 100 100 100 100 100 [B] 111 49 164 29 52 [C] 28 15 64 29 52 Precision (%) 74.77 69.38 60.97 0.0 0.0 Recall (%) 83.0 34.0 100.0 0. 0 0. 0

Table 4. Detection performance of SVM classiers trained with stego images of specic embedding techniques using Binary Similarity Metric(BSM) as features for Dataset 1 with embedding capacity of 4.5 bpp; [A]: Number of actual datamasked image streams; [B]: Number of image streams detected by the steganalyzer; [C]: Number [C ]) [C ]) of false alarms; Precision: ([ B ] ; Recall: ([ B ] . [B] [ A] [A] F5r12 F5r11 Out Out+ Lsb 100 100 100 100 100 [B] 23 66 30 21 121 [C] 23 20 25 20 89 Precision (%) 0.0 69.69 16.66 4.76 26.44 Recall (%) 0.0 46.0 5.0 1.0 32

Table 3. Detection performance of linear classiers trained with stego images of specic embedding techniques using Binary Similarity Metric(BSM) as features for Dataset 2 with embedding capacity of 4.5 bpp; [A]: Number of actual datamasked image streams; [B]: Number of image streams detected by the steganalyzer; [C]: Number [C ]) [C ]) of false alarms; Precision: ([ B ] ; Recall: ([ B ] . [B] [ A] [A] F5r12 F5r11 Out Out+ Lsb 71 71 71 71 71 [B] 89 56 111 43 48 [C] 28 20 40 43 48 Precision (%) 68.54 64.29 63.96 0.0 0.0 Recall (%) 85.92 50.7 100.0 0.0 0.0

Table 5. Detection performance of SVM classiers trained with stego images of specic embedding techniques using Binary Similarity Metric(BSM) as features for Dataset 1 with embedding capacity of 2.8 bpp; [A]: Number of actual datamasked image streams; [B]: Number of image streams detected by the steganalyzer; [C]: Number [C ]) [C ]) ; Recall: ([ B ] . of false alarms; Precision: ([ B ] [B] [ A] [A] F5r12 F5r11 Out Out+ Lsb 100 100 100 100 100 [B] 23 58 29 20 106 [C] 23 20 25 20 89 Precision (%) 0.0 65.52 13.79 0.0 16.04 Recall (%) 0.0 38.0 4.0 0.0 17.0

Table 5 presents the results of a linear classier on dataset 1 with embedding capacity of 2.8 bpp. Compare this with Table 1 which presents the detection results for the same classier and dataset but with higher embedding capacity of 4.5 bpp. Table 6 presents the results of a linear classier on dataset 2 with embedding capacity of 2.8 bpp. Compare this with Table 2 which presents the detection results for the same classier and dataset but with higher embedding capacity of 4.5 bpp. From all the experiments we consistently observe that, the detection rate of classiers decreases as one decreases the embedding capacity of datamasking algorithm. Also, the reduction in detection rate was only marginal. While these experiments were performed with steganalyzers trained with stego images of specic embedding techniques, we also performed an experiment with a generic steganalyzer trained with stego im-

Table 6. Detection performance of SVM classiers trained with stego images of specic embedding techniques using Binary Similarity Metric(BSM) as features for Dataset 2 with embedding capacity of 4.5 bpp; [A]: Number of actual datamasked image streams; [B]: Number of image streams detected by the steganalyzer; [C]: Number [C ]) [C ]) of false alarms; Precision: ([ B ] ; Recall ([ B ] . [B] [ A] [A] F5r12 F5r11 Out Out+ Lsb 71 71 71 71 71 [B] 15 13 23 6 75 [C] 14 9 18 6 66 Precision (%) 6.67 30.77 21.74 0.0 12.00 Recall (%) 1.41 5.63 7.04 0. 0 12.68

ages of different embedding schemes. The detection performance of the generic steganalyzer on the same set of datamasked images was found to be 2.81% at 4.5 bpp and 1.31% at 2.8 bpp. Also, the embedded message length estimate using RS steganal-

302

Radhakrishnan, Kharrazi and Memon

Table 7. Detection performance of SVM classiers trained with stego images of specic embedding techniques using Binary Similarity Metric(BSM) as features for Dataset 2 with embedding capacity of 2.8 bpp; [A]: Number of actual datamasked image streams; [B]: Number of image streams detected by the steganalyzer; [C]: Number [C ]) [C ]) of false alarms; Precision: ([ B ] ; Recall ([ B ] . [B] [ A] [A] F5r12 F5r11 Out Out+ Lsb 71 71 71 71 71 [B] 14 11 22 6 83 [C] 14 9 18 6 66 Precision (%) 0.0 18.18 18.18 0.0 20.48 Recall (%) 0. 0 5.63 5.63 0. 0 23.94

masked images to learn the statistics of the datamasking scheme. In order to verify that we also trained a steganalyzer using datamasked images and have found that both linear & SVM classiers have a precision of above 87% while maintaining a recall rate above 97%.

4.

Conclusions

Table 8. Detection performance of linear classiers trained with stego images of specic embedding techniques using Binary Similarity Metric(BSM) as features for Dataset 1 with embedding capacity of 2.8 bpp; [A]: Number of actual datamasked image streams; [B]: Number of image streams detected by the steganalyzer; [C]: Number [C ]) [C ]) ; Recall ([ B ] . of false alarms; Precision: ([ B ] [B] [ A] [A] F5r12 F5r11 Out Out+ Lsb 100 100 100 100 100 [B] 109 53 164 29 52 [C] 28 15 64 29 52 Precision (%) 74.31 71.7 60.97 0.0 0.0 Recall (%) 81.0 38.0 100.0 0. 0 0. 0

Table 9. Detection performance of linear classiers trained with stego images of specic embedding techniques using Binary Similarity Metric(BSM) as features for Dataset 2 with embedding capacity of 2.8 bpp; [A]: Number of actual datamasked image streams; [B]: Number of image streams detected by the steganalyzer; [C]: Number [C ]) [C ]) of false alarms; Precision: ([ B ] ; Recall ([ B ] . [B] [ A] [A] F5r12 F5r11 Out Out+ Lsb 71 71 71 71 71 [B] 91 44 114 43 48 [C] 28 20 40 43 48 Precision (%) 69.23 54.55 62.28 0.0 0.0 Recall (%) 88.71 33.80 100.0 0.0 0.0

In this paper, we proposed a novel solution to remove randomness from cipher streams in a reversible manner. This would make difcult for an eavesdropper to distinguish encrypted streams from other network streams, thereby, providing covertness to secure channels. We have proposed schemes to generate audio streams and image streams from encrypted streams. Since perceptual tests are not feasible with such large volumes of data, we relax the requirement for perceptual transparency to increase the steganographic streams. However, we maintain certain statistical properties in the datamasking process: the shape of power spectral density in the case of audio datamasking and the shape of prediction error PDF in the case of image of datamasking. Our experimental results show that the proposed datamasking schemes have a different signature than existing steganographic schemes. Even though existing steganalyzers do not perform well to detect datamasked streams, we have also shown that one can train a steganalyzer with datamasked streams to detect them with high accuracy. For the existing steganalyzers, ensuring spectral shape and prediction error PDFs shape was sufcient. However, it remains to be seen if one can devise a framework for steganography which would escape any other higher order statistical analysis. Our experiments seem to suggest that an embedding mechanism which induces similar noise characteristics as the noise that is already present in the image, would be hard to be detected by any statistical analysis.

Acknowledgments We would like to thank Kulesh Shanmugasundaram (CIS department, Polytechnic University, Brooklyn, NY) for many helpful discussions on data masking and network security. We are also grateful to Hamza Ozer (National Research Institute of Electronics and Cryptology, Marmara Research Center, Gebze, Turkey) for

ysis was only 25% of the actual embedded message length. All of these experiments suggests that theproposed datamasking scheme has a different signature from that of existing steganographic schemes. Hence, one needs to train a new steganalyzer with data-

Data Masking: A New Approach for Steganography?

303

helping in evaluating datamasked streams with the audio steganalyzers.

6.

References
1. G.J Simmons, Prisoners Problem and the Subliminal Channel (The), CRYPTO83Advances in Cryptology, Aug. 1984, pp. 5167. 2. N.F Johnson and S. Katzenbeisser, A Survey of Steganographic Techniques, in Information Hiding S. Katzenbeisser and F. Petitcoals (Eds.), Artech House, Norwood, MA, 2000, pp. 4378. 3. N.F. Johnson and S. Jajodia, Steganalysis of Images Created Using Current Steganography Software, in Information Hiding, David Aucsmith (Ed.), LNCS 1525, Springer-Verlag Berlin Heidelberg, 1998, pp. 3247. 4. A. Westeld and A. Ptzmann, Attacks on Steganographic Systems, in Information Hiding, LNCS 1768, Springer-Verlag Heidelberg, 1999, pp. 6176. 5. J. Fridrich, M. Goljan, and R. Du, Reliable Detection of LSB Steganography in Color and Grayscale Images, in Proc. of the 7. 8.

9.

10.

11.

12.

ACM Workshop on Multimedia and Security, Ottawa, CA, Oct., 2001, pp. 2730. I. Avcibas, N. Memon, and B. Sankur, Steganalysis Using Image Quality Metrics, Security and Watermarking of Multimedia Contents, SPIE, San Jose, 2001. Advanced Encryption Standard, FIPS 197, 2001. ISO/IEC 14495-1, ITU Recommendation T.87, Information technologyLossless and Near-Lossless Compression of Continuous-Tone Still Images, 1999. H. Zer, I. Avcibas, B. Sankur, and N. Memon, Steganalysis of Audio Based on Audio Quality Metrics, submitted to SPIE 2004. Elke Franz et al., Steganography Preserving Statistical Properties, Fifth International Workshop on Information Hiding, 2002. S. Lyu and H. Farid, Detecting Hidden Messages Using Higher-Order Statistics and Support Vector Machines, Fifth International Workshop on Information Hiding, 2002. S. Dumitrescu, X. Wu, and Z. Wang, Detection of LSB Steganography via Sample Pair Analysis, IEEE Transactions on Signal Processing, 2003, pp. 19952007.