Abstract

We propose a new watermarking method based on quantization index modulation. A concept of initial data loss is introduced in order to increase capacity of the watermarking channel under high intensity additive white Gaussian noise. According to the concept some samples in predefined positions are ignored even though this produces errors in the initial stage of watermark embedding. The proposed method also exploits a new form of distribution of quantized samples where samples that interpret “0” and “1” have differently shaped probability density functions. Compared to well-known watermarking schemes, this provides an increase of capacity under noise attack and introduces a distinctive feature. Two criteria are proposed that express the feature numerically. The criteria are utilized by a procedure for estimation of a gain factor after possible gain attack. Several state-of-the-art quantization-based watermarking methods were used for comparison on a set of natural grayscale images. The superiority of the proposed method has been confirmed for different types of popular attacks.

1. Introduction

Digital media have a great impact on many aspects of modern society. Some aspects assume that we deal with audio-visual data that relates to a person or an organization. Information about the relation quite often should be preserved. Watermarking approach is to insert the information in the media itself [1]. However, in that case the watermark might be intentional or not altered by the third party. In order to avoid an alteration the watermark needs to be robust [2]. Other characteristics except robustness may also be important. Watermark invisibility and payload are among them. Invisibility is important to assure that the quality of the media does not degrade significantly as a result of watermarking [3]. High data payload might be needed in some applications in order to define many aspects of ownership.

In the field of digital image watermarking (DIW) digital images are used as a media (or host). DIW incorporates many different techniques and one of the most popular among them is quantization index modulation (QIM). Methods that belong to QIM are widely used in blind watermarking where neither original media nor watermark is known to the receiver [4]. For the purpose of evaluating robustness the watermarked image is being attacked and additive white Gaussian noise (AWGN) is the most popular condition for that. Theoretical limit of the channel capacity which is achievable by QIM under AWGN was first derived in [5].

In most cases quantization is implemented to some coefficients rather than to signal samples. In order to obtain coefficients a transform is applied to a host signal. It has been shown that some transforms provide coefficients that are more robust to popular image processing algorithms like JPEG, geometric modifications, and so forth [6, 7].

It is assumed that during quantization each of the original coefficient values belongs to one of equally spaced intervals. Further, inside each interval coefficients to interpret “0” and “1” are selected. The task of quantization is to separate coefficients that represent different bits inside each interval. The separation efficiency influences robustness and invisibility. The result of the separation can be characterized by the size of original interval, distribution of separated samples, and the distortion incurred by the separation.

However, all the known implementations of QIM are far from achieving the capacity limit under AWGN. The simplest scalar realization of QIM is to replace all the coefficient values from a certain interval by a single value equal either to the left or right endpoint depending on a bit of a watermark [8]. Hence, the distribution of quantized samples that represent both “0” and “1” is degenerate (or Dirac). Nevertheless, the capacity of the simplest QIM (further referred to as QIM without “simplest”) is less than 10% of the limit value for the condition when noise and watermark energies are equal. More advanced realization of DC-QIM is to replace each coefficient value from an original interval by a corresponding value taken from one out of two disjoint intervals that are subsets of the original one [9]. A parameter is to control the size of these intervals relatively to the original. The distribution for “0” and “1” in that case is uniform. Parameter is being adjusted depending on noise level in order to maximize capacity. Method DC-QIM is widely used and provides the highest capacity under AWGN among known practical realizations. However, considering AWGN attack only, the most evident gap under high noise intensity is caused by low capacity in comparison with the theoretical limit.

Some other modifications of QIM have emerged over the past years. Forbidden zone data hiding (FZDH) modifies only a fraction (controlled by ) of coefficient values in each interval of original values [10]. Despite the fact that FZDH performs slightly worse than DC-QIM the paper represents a promising idea on how to reduce embedding distortions. Another idea was proposed by the authors of Thresholded Constellation Modulation (TCM) that use two different quantization rules to modify coefficients inside the original interval [11]. Each rule is applied only to samples from a particular subinterval and defines their endpoints. The value of a shift is different for any value from a subinterval according to the first quantization rule. The second rule is to provide an equal shift to all the values from another subinterval. There are two shift directions in order to embed “0” and “1.”

The main advantage of the techniques based on QIM with different kinds of compensation [9–11] is a considerable robustness against AWGN. The limitation is that synchronization is required to extract a watermark. Even minor distortion of a different kind can make embedded information unreadable. The simplest realization of such kind of distortion is a gain attack (GA) which performs a constant scaling of the whole sequence of watermarked coefficients. The scaling factor might be close to 1 and cause very little visual distortion, but it is unknown to the receiver which causes asynchronous extraction. Usually GA is followed by AWGN that complicates retrieval of the watermark [12]. Vulnerability to GA causes one of the most critical gaps for practical implementation of QIM-based methods with compensation.

Many different approaches have been developed to improve robustness against GA of QIM related methods [13]. Most approaches can be classified into two groups where the main idea of the first group is to estimate the unknown factor while the idea of second is to quantize coefficients that are invariant to scaling of original signal.

Estimation of the scaling factor requires modelling. Some feature that is unique for the watermarked and attacked signal might be described by a model [14]. The scaling factor may be included in the model and to be a subject to optimization. An obvious complication is that a process of feature selection is not straightforward. In some cases the feature is created instead of being selected and some permanent data agreed upon between the sender and the receiver is a suitable example. However, such compulsory agreement limits practical implementation of watermarking method. Other possible limitations are low model accuracy or computationally heavy optimization.

For instance, a kind of GA and a constant offset attack followed by AWGN are assumed in [12]. The solution proposed there is to embed a pilot signal and to use Fourier analysis to estimate the gain factor and the offset. However, an obvious disadvantage of the solution is that the precision of estimated parameters is low even for quite long pilot sequence.

The method of recovery after GA and AWGN is proposed in [15]. It uses information about dither sequence and applies maximum likelihood (ML) procedure to estimate the scaling factor. The estimation is based on a model of watermarked and attacked signal. Information about statistical characteristics of original host signal should be either known or guessed in order to define the model. Another limitation of the approach is that it requires exact information about embedding parameter . Computational complexity of the approach is high.

As an opposite for estimation, invariance to GA, in general, requires more complex transform of original signal (e.g., nonlinear,) to obtain coefficients. It is necessary to modify coefficients to embed a watermark. However, a model to estimate distortions of the host is more complex in that case. Distortions should be controlled which limits the choice of the kind of QIM to one that adds less complexity to a model of distortion. This, for example, might result in reducing the number of adjustable parameters of QIM. This is one of the reasons why invariant to GA approaches are more vulnerable to AWGN compared to DC-QIM.

Rational dither modulation (RDM) is one of the most popular watermarking methods invariant to GA [16]. For a particular coefficient, a ratio that depends on a norm of other coefficients is being quantized instead of a coefficient itself. The simplest QIM scheme is utilized in order to quantize the ratio and the performance of RDM under AWGN (without GA) is close to the simplest QIM. Other recent blind watermarking methods robust to GA are proposed in [17–23]. Nevertheless, for GA invariant methods the gap is caused by the reduced capacity under AWGN.

In this paper, we propose our own scalar QIM-based watermarking approach that is beneficial in several aspects. The approach addresses the mentioned gaps in the literature: it both delivers higher capacity under AWGN and recovers after GA. In order to do this, host signal coefficients are separated in a way that the resulting distributions for coefficients that interpret “0” and “1” are different. This distinctive feature is used by a simple yet efficient procedure for estimation of a scaling factor under GA. A concept of initial data loss (IDL) is introduced in order to increase watermark channel capacity under low watermark to noise ratios (WNRs). According to IDL, some fraction of wrong watermark bits is accepted during embedding procedure.

The rest of the paper is organized as follows. In Section 2 we describe our quantization model using formal logic approach and derive some constraints on the parameters of the model. In Section 3 some important watermarking characteristics of the model are evaluated analytically while the following Section 4 contains description for the procedure of recovery after GA as well as experimental results obtained under popular attacks. In Section 5 we discuss in detail experimental conditions and compare the performance of the proposed method with the performance of well-known methods. Section 6 concludes the paper and outlines possible directions for improvement.

2. Quantization Model

In this section we define a new model of quantization. First, it is necessary to show that according to our model the separation of original coefficients is possible and we can embed information. Formal logic approach is used to define dependencies between several conditions that are important for the separation of original coefficients. Separation argument (SA) represents the model in a compact form yet has a clear structure which is sufficient to reason the intuition behind the dependencies. Second, it is necessary to assure conditions when SA is sound.

2.1. Formalization of SA

Symbol will be used to denote a random variable whose domain is the space of original coefficients of a host. A particular realization of will be denoted by . We will further describe our model for values that are in some interval of size . More specifically we will refer to an interval with integer index whose left endpoint is . Such an interval is referred to further as embedding interval. For any we define and will be used to denote a random variable which represents . The length is selected in a way that an appropriate document to watermark ratio (DWR) is guaranteed after the separation. We also assume that is small enough to derive that the distribution of is uniform. A random variable that represents separated coefficients inside th interval is denoted by and its realization is denoted by . Correspondingly, a random variable that represents separated coefficients on the whole real number line is denoted by and its realization is denoted by . Each pair of an original and corresponding quantized belong to the same th embedding interval so that an absolute shift is never larger than .

Let us denote a watermark bit by . Truncated pdfs and are used to describe the distribution of and should be defined prior to quantization. Parameters and represent fractions of IDL for and , respectively. Parameters and represent fractions of the samples where original values are to be modified by a quantizer for and , respectively. It is therefore assumed that the fraction of zeros in a watermark data is and fraction of ones is . Condition always holds. The result of the separation in the th embedding interval depends on , , , , , , and . In other words is defined by quantizer that has the mentioned parameters:

We will use SA to describe the quantizer . Each of logical atoms , and represents some condition which is either true or false: For example, is true if and only if “” and is not classified for IDL. We formalize SA in the following way: It can be seen that SA is valid. The conclusion of SA states that the separation of coefficient values inside th embedding interval is possible which means that the proposed model is suitable for information embedding. Furthermore, each premise represents an important dependency between input and output of the quantizer and we require that each premise is indeed true. Hence, it is necessary to enforce soundness for SA.

The intuition behind SA can be explained in the following way. Initially samples with labels “” and “” are not separated in the dimension of inside the mentioned th embedding interval. In order to separate them we shift those with “” to the left and those with “” to the right. If so, shift to the right for “” or shift to the left for “” is not acceptable because it would introduce distortion and on the other hand worsen separation between “0” and “1.” Therefore for formula is true and for formula is true.

Another consideration is that for any two with the same bit value we infer that quantization in a way that implies less distortion than if . Saving the order we preserve cumulative distribution in respect to the order. Quantized samples that interpret “0” are distributed according to pdf ; samples that interpret “1” are distributed according to pdf . Therefore or is true if or is true, respectively.

And, lastly, the condition for IDL is and it is the case when is not modified and therefore .

An illustration of an example where SA is sound is given in Figure 1. Two positions of original values are shown on the lower part of Figure 1. Condition is satisfied for the first original value and condition is satisfied for the second. Two positions of the modified values are shown on the upper part of Figure 1. After the separation the modified values satisfy conditions and , respectively. The areas of green segments on the lower and the upper parts of Figure 1 are equal. The areas of blue segments are also equal. As it can be seen on the upper part of Figure 1, the distribution of separated coefficients in th embedding interval depends on and .

Figure 1 

Illustration of the process of separation.

Parameters of the pdfs and need to be specified in order to prove soundness for the whole range of in the th interval. In addition formulas (7) and (8) need to be rearranged in order to express in a suitable way for the quantization form.

We propose such and that in general there is no line of symmetry which can separate them inside embedding interval. This feature will provide easier recovery after GA. It is necessary to emphasize that the proposed functions and only describe distributions for fractions and , respectively (e.g., without taking into account fractions of IDL). We introduce parameters and to define both pdfs and , where , as shown in Figure 2(a). As can be seen the density is zero in the subinterval which separates “0” from “1.” In Figure 2(b) we can see the distribution of the quantized coefficients outside th embedding interval as well.

(a)

(b)

(a)
(b)

Figure 2 

Distribution of the quantized coefficients: (a) inside th embedding interval; (b) in five consecutive intervals.

Namely, the proposed truncated pdfs are a linear function and a constant: The samples that belong to IDL fraction are distributed according to pdfs and :

2.2. Soundness Conditions for SA

The soundness of SA is guaranteed if it is possible to satisfy or when or is true, respectively. The requirement to satisfy or imposes some constraints on , and . Let us find those constraints.

We start from defining parameters of and using property of pdf: It is easy to derive from (14) that According to (4), (5), and (7) condition is satisfied if and only if for all Using (10) and the fact we can derive The latter inequality should be true for all which means For our particular application we chose ; therefore and we are using the value in our method.

Using (13) we can conclude that

Functions and can be fully defined now. Let us find dependencies that connect and with , and . Taking into account that in our realization we can derive from (20) that

According to (4), (6), and (8) condition is satisfied if and only if Using (15) we find that

In the experiment section of the paper the goal is to find the highest capacity for a given WNR. Different values of the parameters need to be checked for that purpose. Preserving (15) and (19)–(21), (23) would guarantee soundness of SA and avoidance of using parameters’ combinations that are not efficient for watermarking. This can reduce required computations.

2.3. Embedding Equations

For the proposed pdfs we can now define as a function of , which is the main task of the quantizer . Let us consider conditions , separately as it is never the case when both conditions are true. We will denote in case of by , but in case of the notation will be used.

From (7), (10), and it is clear that Taking into account that we derive From (8), (11), and (15) we can find that

According to (26), the values of quantized coefficients are linearly dependent on original values while according to (25) the dependency is nonlinear. Different character of dependency between quantized and original values for “0” and “1” is one of the key features of our approach. This differentiates the proposed watermarking method from the methods previously described in the literature [10–12].

3. Characteristics of Quantization Model

The model was proposed in the previous section. It was shown that it is suitable for coefficient separation and the conditions necessary for soundness of SA were defined. In this section we focus on efficiency of separation. The main characteristic that can be estimated analytically is the watermark channel capacity under AWGN. It is required to calculate such characteristic for different WNRs. First, we express WNR in terms of parameters of the quantization scheme. Second, we express error rates in terms of parameters of the quantization scheme. This makes it possible to include WNR in the expression for error rates (and capacity).

3.1. Estimation of Quantization Distortions

The variance is the only parameter of AWGN attack and is defined as where is a watermark energy. Alternatively, can be seen as a distortion of a host signal, induced by the quantization. Let us define .

For the matter of convenience of the experiment it is better to use a single parameter (control parameter) that can be adjusted in order to provide the desired value of . While defining we choose to be the control parameter and collect it in the expression for . The total distortion is a sum of distortions and caused by two types of shifts that are and , respectively. The first distortion component is defined as Proceeding further and using (10) we can derive that However, it is clear from (19)-(20) that both parameters and depend on . In order to collect we introduce two independent of parameters and . This brings us to

The second distortion component is defined as Using (11), (15), and integrating in (31) we obtain

The total quantization distortion can be expressed in terms of , and :

For any combination of , and the required value of is defined using (27) and (33) as

3.2. Estimation of Error Rates

Bit error rate (BER) and channel capacity can be calculated without simulation of watermark embedding procedure. It is important that the kind of threshold used to distinguish between “0” and “1” is suitable for analytic estimations. Further we assume that the position of the threshold remains permanent after watermark is embedded and does not depend on attack parameters. In Figure 2(b) the position of the threshold is Th for intervals numbered ,  . For the intervals numbered the position of the threshold is .

The absolute value of quantized sample in any interval is . We use for a sample that is distorted by noise. Hence, interprets “0” or “1” depending on belonging to or , respectively:

There are two cases when errors occur in non-IDL samples. An error in “0” is incurred by a noise if and only if the both following conditions are true:

An error in “1” occurs if and only if the following is true:

Two cases when errors occur in IDL samples can be presented with the following conditions for “0” and “1,” respectively:

The pdf of AWGN with variance can be represented in terms of and as . In general we can estimate error rates for an interval with any integer index . For that purpose we use generalized notations , , , and for pdfs of quantized samples in any interval. For example, for even pdf ; for odd pdf . We denote interval by . Then, the error rates for quantized samples in can be defined as

Now we can show that BER₀ and BER₁ can be calculated according to (41) for any chosen interval. For that purpose it is enough to demonstrate that any component in (41) remains the same for every interval. For example, we state that for any .

Let us first assume , . Then, , . However, it is also clear from (36) that . Hence, and we prove the statement.

Now let us assume , . Then, , . For the matter of convenience we accept that for some . Therefore . Also . The property of pdf of AWGN provides that and, consequently, Using the latest equation we derive that and we prove the statement.

4. Experimental Results

In this section we describe conditions, procedure, and results of two different kinds of experiments based on analytic estimation of capacity as well as simulations. The preferred index of attack severity is WNR (indexes and quality of JPEG compression are also used). For a given set of embedding parameters, the error rates and capacity are estimated differently using different models suitable for each kind of experiment. However, for both kinds of experiment, the maximum capacity for a given level of attack severity is found by using brute force search in the space of all adjustable parameters.

4.1. Analytic Estimation of Watermarking Performance under AWGN

In this subsection of our experiment we use . Parameters , and are subjects to constraints (21), (23), , and and the simulations are repeated for each new value of . Then, the length of embedding interval is calculated according to (34). Error rates are calculated according to (41).

We use two variants of the proposed quantization scheme with adjustable parameters: nonsymmetric QIM (NS-QIM) and nonsymmetric QIM with IDL (NS-QIM-IDL). Such a decision can be explained by a consideration that IDL is acceptable for some application, but other applications may require all the watermark data to be embedded correctly.

In Figure 3 the plots for channel capacity toward WNR are shown for two variants of the proposed method as well as DC-QIM and QIM [9]. The permanent thresholding is applied to NS-QIM and NS-QIM-IDL. As a reference, Costa theoretical limit (CTL) [5] is plotted in Figure 3:

Figure 3 

Analytic-based estimation of capacity under AWGN.

Capacity is calculated analytically according to the description provided in the literature for DC-QIM and QIM. During the estimation, the subsets and were used instead of and :

Therefore, for such estimation we assume that quantized coefficients from the th interval after AWGN are distributed only inside . The assumption is a compromise between computational complexity and the fidelity of the result.

As can be seen from Figure 3 both variants of the proposed method perform better than DC-QIM for WNR values less than −2 dB and, obviously, much higher capacity provided by DC-QIM-IDL is compared to the other methods in that range. Taking into account that DC-QIM provides the highest capacity under AWGN compared to the other known in the literature methods [12, 19], newly proposed method DC-QIM-IDL fills an important gap. Reasonably, the demonstrated superiority is mostly due to IDL.

4.2. Watermarking Performance in Simulation Based Experiments without GA

The advantage of analytic estimation of error rates according to (41) is that the stage of watermark embedding can be omitted and host signal is not required. The practical limitation of the approach is that and are just subsets of and , respectively. Other disadvantages are that estimation might become even more complex in case the threshold position is optimized depending on the level of noise; only rates for AWGN can be estimated, but there are other kinds of popular attacks [24]. Therefore in this subsection we will also simulate watermarking experiments using real host signals.

4.2.1. Conditions for Watermark Embedding and Extraction

In case of experiments with real signals the parameters of the proposed watermarking scheme must satisfy some other constraints instead of (34). However, constraints (21), (23), , and remain the same as in the analytic based experiment.

Some lower limit of DWR has to be satisfied for watermarked host, which assures acceptable visual quality. DWR is calculated according to where is the variance of the host.

Therefore, using (33) the equation for in that case is

In contrast to analytic based experiment, should be adjusted for different severity of the attack and is defined as

After watermark is embedded and AWGN with is introduced we perform extraction and calculate channel capacity.

A variant NSC-QIM with constant (nonadjustable) parameters is also used in some experiments. The intention to adjust the parameters in order to maximize capacity is natural. However, maximization requires information about WNR to be known before watermark embedding and transmission. In some application areas level of noise (or severity of an attack) might change over time or remain unknown. Therefore watermark should be embedded with some constant set of parameters depending on expected WNR.

Different positions of the threshold can be used to extract a watermark. An optimal position of the threshold is not obvious. Placing the threshold in the middle of the interval might be inefficient because the distribution of quantized samples inside embedding interval is nonsymmetric. Two kinds of thresholding are proposed: permanent and nonpermanent. The permanent position is for the intervals with numbers , . The name “permanent” is because cannot be changed after embedding. Its position depends only on , and and does not depend on the parameters of attack.

The nonpermanent position of is the median of the distribution inside each interval. Nonpermanent position may depend on the type and severity of a noise. The advantage of nonpermanent is that extraction of a watermark can be done without information about and .

4.2.2. Watermarking Performance for AWGN and JPEG Attacks without GA

The performance of the proposed method was evaluated using real host signals. For that purpose we used 87 natural grayscale images with resolution 512 × 512. Each bit of a watermark was embedded by quantizing the first singular value of SVD of 4 × 4 block. This kind of transform is quite popular in digital image watermarking and the chosen block size provides a good tradeoff between watermark data payload and robustness [7, 25]. The value of DWR was 28 dB. An attack of AWGN was then applied to each watermarked image. The resulting capacity toward noise variance is plotted for different methods in Figure 4.

Figure 4 

Capacity under AWGN for natural grayscale images.

It can be seen that the resulting capacity after AWGN attack is the highest for NS-QIM. The other two methods whose performance is quite close to NS-QIM are DC-QIM and FZDH. Compared to DC-QIM the advantage is more obvious for higher variance. However, for moderate variance the advantage is more obvious compared to FZDH.

Methods QIM and RDM do not have parameters that can be adjusted to different variance. Under some circumstances adjustment is not feasible for NS-QIM as well. We have chosen constant parameters and for NSC-QIM in order to provide a fair comparison with QIM and RDM. The plots for NSC-QIM, QIM and RDM are marked by squares, triangles, and crosses, respectively, in Figure 4. As can be seen, NSC-QIM performs considerably better than QIM and RDM and the advantage is especially noticeable for higher noise variance.

Other image processing techniques except additive noise are able to destroy a watermark and one of them is JPEG compression which is quite popular. The capacity of the proposed watermarking method was also compared with other methods and the procedure of embedding was the same as in AWGN case. However, this time JPEG compression with different levels of quality was considered as an attack. The results are plotted in Figure 5.

Figure 5 

Capacity under JPEG for natural grayscale images.

According to the plots in Figure 5, the performance of NS-QIM in general is very close to that of DC-QIM but is slightly worse for low factor. The methods FZDH and TCM provide lower capacity than NS-QIM and DC-QIM but in general are quite close to them. The worst performance is demonstrated by QIM and RDM and the disadvantage is especially noticeable for low . For NSC-QIM with and the performance is considerably better than that for QIM and RDM under low Q but is worse for higher quality of JPEG compression.

4.3. Procedure for GA Recovery

It has been demonstrated that for some popular types of attack the performance of NS-QIM is comparable or better than that of DC-QIM. The mentioned DC-QIM is considered to be one of the best quantization methods for watermarking, but it is extremely vulnerable to GA. On the other hand the performance of RDM is not as good under AWGN and JPEG attacks and is comparable to that of QIM. In this subsection, we propose a procedure for GA recovery in order to fill an important gap in the literature and introduce a watermarking method that provides high efficiency under AWGN as well as GA. The procedure utilizes features that are unique for the proposed approach and have not been discussed in the field of watermarking before.

We are proposing several criteria that will be used by the procedure to provide robustness against GA for NS-QIM. The criteria exploit nonsymmetric distribution inside embedding interval and help to recover a watermarked signal after the attack. It is presumed that a constant gain factor is applied to the watermarked signal (followed by AWGN) and the task is either to estimate the factor or the resulting length of embedding interval.

Let us denote the actual gain factor by and our guess about it by . The length of the embedding interval (which is optimal for watermark extraction) is modified as a result of GA and is denoted by . Our guess about is .

The core of the procedure of recovery after GA is the following. For each particular value , noisy quantized samples are being projected on a single embedding interval:

One of the following criteria is being applied to the random variable :