Forensic Detection Using Bit-Planes Slicing of Median Filtering Image

For the detection of median filtering (MF) forensics, this paper proposes the feature vector extracted from the bit-planes slicing of the forged image. The assembled feature vector is trained in a support vector machine (SVM) classifier for the MF detection (MFD) of the forged images. The performance of the proposed MFD scheme is measured with several types of forged images: unaltered, Gaussian filtering (<inline-formula> <tex-math notation="LaTeX">$3\times 3$ </tex-math></inline-formula>), averaging filtering (<inline-formula> <tex-math notation="LaTeX">$3\times 3$ </tex-math></inline-formula>), downscaling (0.9), upscaling (1.1), and post-frame-up, respectively, in a block size <inline-formula> <tex-math notation="LaTeX">$32\times 32$ </tex-math></inline-formula> and <inline-formula> <tex-math notation="LaTeX">$64\times 64$ </tex-math></inline-formula> pixels. Subsequently, in experimental items, a classification ratio, Area Under the Curve (AUC), <inline-formula> <tex-math notation="LaTeX">${P} _{\mathrm {TP}}$ </tex-math></inline-formula> at <inline-formula> <tex-math notation="LaTeX">${P} _{\mathrm {FP}} =0.01$ </tex-math></inline-formula>, and <italic>Pe</italic> (a minimum average decision error) are estimated. The result in terms of AUC shows that the estimation of the proposed MFD scheme is graded as ‘ <italic>Excellent</italic> (<inline-formula> <tex-math notation="LaTeX">${A}$ </tex-math></inline-formula>)’.


I. INTRODUCTION
In the image manipulation, the alterations are using filtering, averaging, rotating, and up-down scaling in the forgery methods.Many forgers have particularly preferred MF because an image has the characteristics of nonlinear filtering based on order statistics.Therefore, many works have been conducted for MFD [1]- [3].
For the forensic detection of forged images in this research area, the edge-corner [4] and, the autoregressive (AR) model of median filter residual (MFR AR) [5] used.The MFR means the difference value about the original image and its MF image.In [5], AR coefficients obtained as the feature vector of the MFD to analyze the MFR.
For the image forensics detection, the residual image is used frequently, which includes the features for forensic detection.By doing so, the detection can be operated efficiently with modest pixel transformation.
Yang et al. [6] used a two-dimensional autoregressive (2D-AR) model [7] to characterize the residuals of the filtered images for the MFD.This attempt seemed to have improved the low performance of the MFR AR with the combined more high-dimensional feature set, but only the The associate editor coordinating the review of this manuscript and approving it for publication was Guitao Cao.correlation coefficients of the horizontal and the vertical direction, respectively in an image was used.
In this paper, a new MFD scheme is proposed, in which the three kinds of the feature vector extracted from the two bit-planes and their residual of the forged image, respectively.
Be comprehended in this paper; the main contribution is as follows: 1) The feature vector extracted from the bit-planes in an image, and then which is treated as a bandpass filter.
2) The computing of the pass filter, low-pass, the bandpass and high-pass filter on an image is operated by selecting the corresponding the bit-plane.
3) The results of the experiment are mainly compared to the MFR AR and the MFR 2D-AR methods.4) The proposed scheme applied to the re-tampered actual cut-paste images: {altered, double altered, rotate and add noise}, and confirmed the excellent detection of median filtering in the various of re-tampered images.As follows: in Section 2, the bit-planes slicing and the autoregressive model used in the image forensics briefly descriptively.In Section 3, the feature vector for the proposed MFD algorithm extracted from the AR model of the bit-planes and the residual.The experimental results of the proposed MFD scheme are measured, and the resultant performance of the evaluation compares with the prior work in Section 4. Lastly, the conclusion is drawn in Section 5.

II. RELATED THEORIES A. BIT-PLANES SLICING OF IMAGE
In image processing, an image could be sliced into bit-planes.Many image processing tools can split an image into the bitplanes.Also, an image is digitally represented in terms of pixels, which can be signified a bit.The binary formats for those pixel values are an 8-bit representation.Plane 1 and 8 contain the lowest and the highest order bit of all the pixels in the image, respectively.This bit-planes slicing is a method to be processed in a spacial domain of the image.Fig. 1 shows the bit-plane 1 to 8 of BOWS2 image.It considers plane 1 is a high pass filter, and plane 8 is a low pass filter, respectively.

B. AUTOREGRESSIVE MODEL
In signal processing and statistics, an autoregressive (AR) model is a described a type of stochastic process.It used to describe the specific time-varying process in nature and economics, etc.The AR model designates that the output variable depends on linearity, its previous values, and an imperfectly predictable in term of a stochastic.The model is to be in the form of a stochastic difference equation.
The notation AR (p) indicates an AR model of the order p.Subsequently, the AR (p) is defined as where ϕ l , . . ., ϕ p are the parameters, c is a constant, and e t s white noise of the AR model.Using the backshift operator B, (1) can be equivalently formed as now, the summation term in (2) move to the left side, formed as (3) by polynomial notation Some parameter limitations are required for the model to dwell wide-sense stationary.For example, in (1), the AR model with |ϕ 1 | ≥ 1 are not stationary.More typically, for an AR (p) model to be wide-sense stationary, the roots of the polynomial z p − p l=1 ϕ l z p=l must be placed in the unit circle, i.e., each (complex) root z l must satisfy |z l | < 1.
Kang et al. [5] and Rhee et al. [4], [12] applied the AR model to extract the feature vector of an image.An image I has horizontal and vertical line components h and v, respectively.The AR coefficients are calculated as follows where k is the AR order number, k range is 1 ≤ k ≤ p, and p is the maximum order number.In [4] and [5], the ten AR coefficients to be regarded as the feature vector for the forensic image detection.Again, the AR coefficients generate the dissimilarity image by where ε (r) (m, n) and ε (c) (m, n) are the prediction errors in a horizontal and vertical direction, respectively.q is the neighboring range of (m, n).
Similarly, the 2D-AR model of median-filtered residual is formally defined as where (i, j) is the neighboring range of (m, n), and (p, q) is the maximum order number in the horizontal and the vertical direction, respectively.

III. PROPOSED DETECTION ALGORITHM OF MEDIAN FILTERING FORENSICS
In this section, the feature vector for the proposed MF forensic detection scheme is composed of 30 dimensionalities.Table 1 reports the notations, which utilized in the proposed MFD algorithm.Also, the block diagram of the proposed MFD is shown in Fig. 2, and the scheme is described in Fig. 3.When a suspicious image was given as a query image, which is sliced

IV. EXPERIMENTAL RESULT
First, it describes the adopted experimental methodology in this section.Second, the performance results of the implemented MFD classifier are compared to [5], [6] to estimate the efficiency of the proposed MFD scheme.Lastly, to prove the classifying effectiveness, the implemented MFD detector applied to the actual forged (Cut-Paste) image.Fig. 4 shows the BOWS2 [8] and UCID [9] example images, and their extracted feature vector of the proposed MFD scheme is depicted in Fig. 5 where Unaltered (Original), GAU3 (Gaussian filtering: window size (3 × 3)), MF3 (Median filtering: window size (3×3)), AVE3 (Averaging filtering: window size (3 × 3)), and JPG90 (QF = 90), respectively.

A. EXPERIMENTAL METHODOLOGY
Input the assembled 30-dim.feature vector into a SVM classifier to train of the MF classification.C-SVM with Gaussian kernel [10] is adopted as the classifier.
Furthermore, the train with four-fold cross-validation in a SVM classifier has a conjunctive grid probe in the multiplication grid.The best parameters of C and γ as The probing step size of (m, n) is 0.25 used to achieve the classifier model of the train set, which similar in meaning to [1], [4], [5], and the experiment adopts the image databases: The BOWS2 and the UCID that consist of 10,000 uncompressed gray images of size 512 × 512, and 1,338 uncompressed color images of size 384 × 512 or 512 × 384, respectively.Under the necessity of the image databases, the color images converted to 8-bit grayscale.

B. IMPLEMENTATION OF SVM CLASSIFICATION FOR THE PROPOSED MFD SCHEME
The feature vector of the UCID images of the proposed MFD scheme is trained in the SVM with (10) and (11).In Fig. 6, the three kinds of the trained feature vector: the positive vector (red color) {MF3, MF5, MF35}, the negative vector (cyan color) {Unaltered, AVE3, JPG90} and support vector (small black circle) are shown.
In here, the first and second feature vectors were plotted, and it could be confirmed that the positive and the negative vector are classified, and little the support vector (uncertain vector) where MF35 means that composites MF3 and MF5.For the estimation in the experiment, the proposed MFD scheme has four kind experiment items: a classification ratio, AUC (Area Under ROC Curve) by the P TP (True Positive rate: Sensitivity) and P FP (False Negative rate: 1-Specificity), P TP at P FP = 0.01, and a minimum average decision error (Pe) which has lower, the reliability of the MFD is higher.
In succession, the trained SVM classifier is used to perform the classification on the test set.Among the BOWS2 and the UCID 11,388 images, randomly selected 80% images are used as the train set, and the rest of the 20% images are the test set.
As a result, in Table 2 the AUCs of the proposed MFD scheme approach to 1. Also, P TP at P FP = 0.01 is almost higher, and Pe is almost lower, respectively, compared  to [5], [6] (good results are presented the bold type of shade cell).Notably, in the test group image A, the proposed MFD scheme has excellent performance idealized.Also, in the test group B and C, MF+DN and MF+JPEG images have excellent performance too.
In the aspect of the whole experiment item, it confirms that outstanding to compare with [5], [6].However, the AUCs are 0.9 over.Thus, the evaluation [11] of the proposed MFD scheme graded as 'Excellent (A)'.

C. MFD CLASSIFIER APPLIED TO CUT-PASTE FORGED IMAGE
In Section 4.2, the SVM classifier is fulfilled with the proposed MFD scheme.It needs to estimate the efficiency of the trained SVM classifier.
So, in Fig. 10, the forged image produced using the images of the BOWS2 database [8], which is combined with the cut (gate: unaltered) and paste (face: median filtered) area.
In     respectively.In each three kind result images, the first, third and fifth are evaluated at 32 × 32 block size, and the second, fourth and sixth columns are evaluated at 64 × 64 block size, respectively.methods and the proposed MFD scheme, respectively.Also, the comparison of the AUCs between both schemes is presented in Table 3.

TABLE 3.
Comparison of AUC between the MFR AR [5], the MFR 2D-AR [6] methods and the proposed MFD scheme in Fig. 11.

V. CONCLUSION
In this paper, a new median filtering detection classifier is proposed.It used the autoregressive model as a method to extract the feature vector.The constructed feature vector as follows; the first one is extracted from the bit-plane 1, the second one is extracted from the bit-plane 8 of the query image, and the last one has extracted the residual between the bit-planes 1 and 8, respectively.Because the residual has all characteristics of the bit-planes 2 to 7, it is processed with one operation without redundancy operation.That way, it was able to process the filter components of low-, high-and band-pass in the image at once.
To the best appreciation, the results are through the key to the median filtering detection.So this scheme works in the assist research area of MF image forensics.
Finally, the proposed MFD scheme can put in a variety solution of legal problems, like the other MFD schemes.

FIGURE 2 .
FIGURE 2. Block diagram of the proposed MFD scheme.

FIGURE 5 .
FIGURE 5. Feature vector of the test image example.

FIGURE 6 .
FIGURE 6. Distribution of the positive and the negative feature vector and the support vector of the BOWS2 and UCID image DB by the proposed MFD scheme.

FIGURE 7 .
FIGURE 7. ROC curves of MF versus test image groups by the MFR AR [5] method.

FIGURE 8 .
FIGURE 8. ROC curves of MF versus test image groups by the MFR 2D-AR [6] method.

FIGURE 9 .
FIGURE 9. ROC curves of MF versus test image groups by the proposed MFD scheme.

FIGURE 11 .
FIGURE 11.Re-tampering images of the Cut-Paste forged example.

Fig. 11 ,
the Cut-Paste forged image re-tampered four kinds: (a) non re-tamper, (b) JPG compression QF = 90, (c) counter-clockwise rotated five degrees, and (d) added salt and pepper noise 5%.Now, the example of actual Cut-Paste images in Fig. 11 applies to the trained SVM classifier in Section 3. As a result, Fig. 12, 13 and 14 shows the three kinds detection images by the MFR AR [5], MFR 2-D AR [6] and the proposed MFD scheme: first and second column is the cut-paste scoring pattern image, third and fourth column is the scoring classification by the scoring pattern, and fifth and sixth are the binary classification by score value 0.

FIGURE 14 .
FIGURE 14. Paste area detection of the proposed scheme.The fulfilled with MFD classifier applied to each image in Fig.11with a 32 × 32, and a 64 × 64 block size, respectively.The score value of the classification between the

FIGURE 15 .
FIGURE 15.ROC curves of the forged images in Fig. 12 by the MFR AR [5] method.

FIGURE 16 .
FIGURE 16.ROC curves of the forged images in Fig. 12 by the MFR 2D-AR [6] method.

FIGURE 17 .
FIGURE 17. ROC curves of the forged images in Fig. 12 by the proposed scheme.