Effective splicing localization based on noise level inconsistencies

Abstract

In the era of digital technology, it becomes easy to share photographs and videos using smartphones and social networking sites to their loved ones. On the other hand, many photo editing tools evolved to make it effortless to alter multimedia content. It makes people accustomed to modifying their photographs or videos either for fun or extracting attention from others. This altering brings a questionable validity and integrity to the kind of multimedia content shared over the internet when used as evidence in Journalism and Court of Law. In multimedia forensics, intense research work is underway over the past two decades to bring trustworthiness to the multimedia content. This paper proposes an efficient way of identifying the manipulated region based on Noise Level inconsistencies of spliced mage. The spliced image segmented into irregular objects and extracts the noise features in both pixel and residual domains. The manipulated region is then exposed based on the cosine similarity of noise levels among pairs of individual objects. The experimental results reveal the effectiveness of the proposed method over other state-of-art methods.

Keywords

Splicing localization object segmentation residual images noise level inconsistencies cosine dissimilarity

1 Introduction

Nowadays, people exchange their ideas, monuments through photographs or videos through social networking sites with their loved ones. Having widely available photo editing tools such as Photoshop or Corel Draw, people easily modify multimedia content either for fun or to get attention from others. One can be an expert to alter in a way that it cannot be traceable visually. When used as documentary proof in Journalism, in the court of law, the pictures play a significant role in supporting or disapproving a particular argument or removing any doubt [6]. This altering brings many security challenges in the trustworthiness of content shared over the internet [35]. The government initiative towards digitization also brings many challenges in securing digital data [34]. These security challenges bring an urgency to find effective and efficient forensic tools capable of detecting malicious attacks and bringing trustworthiness to multimedia content. Image Forensics, a Multimedia Forensics division, aims to explore techniques and tools to detect manipulation attacks on images [10]. In traditional effective forensic methods like watermarking, an authentication code is embedded by some means into the original image, and the same is used for verification [29, 31]. Contrary to these, a blind or passive forgery detection approach assess image authenticity with no external clue [1]. These techniques assume that images took from various cameras or undergone different processing operations introduce different inherent patterns in the targeted image [12]. These underlying patterns were usually consistent in the original image, and the manipulated image will be incoherent. These intrinsic statistics on inconsistencies can be used as a forensic feature to detect manipulation attacks [9]. Different manipulation attacks involve copy-move, splicing, resampling, etc., to tamper digital images. To detect such forgery attacks, the techniques extracts predetermined forensic features from image dataset and apply classification to train the model and evaluate validity of an image from the malicious content attack [19 , 36]. In recent years, the researchers focus not only on tampering detection but also on localizing manipulated region [37]. A lot of research work carried out in the field of detecting copy-move and splicing attacks and there is a need to improve the robustness of localizing splicing attack because of its complex nature. In general, the splicing attack involves two different images; thereby, the area affected will be different from the whole image, which is a fundamental key factor in localizing tampered region [27].

1.1 Related work

Many algorithms in literature introduce some specific operations such as blurring or median filtering in a spliced region, to reveal tampering attacks [16]. Some techniques use fingerprints of the original image, such as Photo Response Non-Uniformity (PRNU), which does not require any assumption but requires the camera model to capture the image [20]. In their works [3 –5] the authors considered the fact that the amount of intrinsic noise in the whole image is generally consistent. Whenever any splicing attack occurs, there are inconsistencies due to different camera models or various processing operations. It is a significant factor in localizing the manipulated area. In [3], the authors detect small spliced regions based on local noise level inconsistencies obtained through wavelet coefficients with non-overlapping blocks. Based on each block noise level homogeneity, the image is segmented into several inhomogeneous sub-regions to expose the spliced region. The same applied to static scene video [1], where Camera Response Function(CRF) measures for each frame and local variance obtains from a noise level function. In [4 , 17], an image is segmented into non-overlapping blocks of equal size, and local noise variance is estimated based on positive kurtosis in the band-pass region. In their work [5], the authors used Principal Component Analysis (PCA) to estimate noise-level for each block. They used two block sizes for finer localization. The disadvantage of block-based methods is that the rate of false alarms is high. To reduce it, [7] used superpixel iterative clustering (SLIC) for irregular segmentation of image rather than block-based. The authors estimated the noise levels on each superpixel at different scales in a residual domain. They derived a noise level function from brightness, standard deviation among superpixels and using k-means clustering expos spliced superpixels. The technique is robust to find more than one suspicious region but multi-scaling takes more computational time. In their work [18], authors extracted the regular shape of each superpixel and estimated noise level based on PCA and noise distribution characteristics. They used Fuzzy means clustering to separate the suspicious superpixels from the original one. In color image steganalysis, residual features extracted from each color channel plays a significant role to improve the performance. In their work [8] the authors obtained noise features from color channel differences and estimated noise level for each superpixel. They used Farthest Distributed Centroids Clustering (FDCC) for localizing spliced superpixels. The technique achieved exceptional localization over other superpixel segmentation methods. Recently, Deep Learning techniques have shown better performance on splicing localization. In [33], the authors proposed a fully convolutional neural network (FCN) method for identifying manipulated regions in synthesized images. This method often makes the image smooth, thereby ignoring small objects. To improve its performance in [21] included a region proposal network (RPN) to the FCN and made the whole network an end-to-end learning system. In [32], the authors propose localization architecture that uses resampling features to capture artifacts through Long short-term memory (LSTM) network and an encoder network designed to differentiate tampered regions from non-tampered. The decoder network in the architecture learns features to localize the tampered region. The final layer, known as softmax, is used to understand the network parameters through the back-propagation algorithm from ground truth masks. The advantage is that this technique is capable of localizing pixel level with high precision. In [24] proposed a two-layer deep convolutional neural network (DNN) in which the first layer learns local residual features. The residual features will be further block pooled with SVM to obtain the final discriminative features for splicing detection. A conditional random field (CRF) is incorporated to localize the tampered region. The method is robust even to JPEG compressed images. These deep learning-based techniques improve accuracy but require training on large labeled databases and high computational complexity. The network extracts high-level visual features and neglects low-level features, which can be sources for forensic cues. To overcome the authors of [28] propose a neural network to learn low-level forensic features capable of detecting splicing forgery trained on a small splicing dataset. The technique is capable of detecting splicing forgery in realistic scenes. This paper proposes a forensic technique that can use low-level features to localize the tampered region from a single image without any prior knowledge of the manipulated attack. Unlike other conventional or deep learning techniques that learn features from the whole picture, we proposed an efficient statistical model that extracts noise level features from individual objects in pixel, residual domains to localize tampered regions.

1.2 Our contribution

The main contribution of the proposed work is: i) An object-based segmentation is used to extract the features. ii) The noise features are estimated from individual objects to reduce computational time. iii) A localization algorithm is proposed to expose splicing attack using cosine similarity among pairs of individual objects of a spliced image. vi) It is the first time to use object segmentation to localize tampered regions in noised based localization techniques The rest of the paper organizes in the following way: Section 2 describes the proposed image splicing localization method. In Section 3, the experimental results and performance analysis are present. Finally, the paper ends with conclusions in section 4.

2 Proposed method

The primary focus of this work is the localization of a spliced region. The proposed work introduces object-level segmentation [30], to improve accuracy and efficiency in localizing the tampered region. The work carried out in three levels, as shown in Fig. 1, namely - i) segment the spliced image into irregular objects, ii) estimate noise levels of each object in pixel and residual domains, iii) estimate pairwise dissimilarity among pairs of objects and localize spliced region.

Fig. 1

The Proposed Frame Work.

2.1 Noise level estimation

In literature, many state-of-art algorithms exist in the field of noise estimation of still images. The works [11, 15] used Principal Component Analysis (PCA) and are considered as standard techniques. However, both approaches underestimate noise variance by taking the smallest eigenvalue of covariance from selected low-rank paths. To improve noise level estimation, in [13] the authors suggested a method based on assumptions that patches taken from noise-free image lie in a low-dimensional subspace that can be learned from PCA rather than spread uniformly across the image.

2.1.1 Noise estimation in pixel domain

The noise estimation objective is to isolate the manipulated region from the original one, so this work uses the method [13] for noise level estimation in the pixel domain.

The noisy image from the original image defined as

$I (x, y) = J (x, y) + η (x, y)$ (1) where J(x,y) is an original noise-free image. I(x,y) be the noisy image of size M X N after adding Gaussian noise η (0, σ²) with zero mean and σ² variance. The image I(x,y) can be decomposed into different patches in the set X_s of size d x d and rearranged as a vector to simplify the performance.

In their work [13], it assumes that clean patches lie in m-dimensional (m - a positive integer) subspace, so the noisy image of each patch x_t in the set X_s redefined as

$x_{t} = {Ay}_{t} + η_{t}$ (2) where A is the m-dimensional subspace with constraint A^TA = I and y_t is the projection point of x_t on the subspace spanned by A. In the eq. 2, to infer the linear model apply PCA where A contains m eigenvectors with m as largest eigenvalues of the covariance matrix $Σ_{x} = \frac{1}{s} Σ_{t = 1}^{s} (x_{t} - μ) (x_{t} - μ)^{T}$ with $μ = \frac{1}{s} Σ_{t = 1}^{s} x_{t}$ .

Define a rotation matrix R = [A, U] with a condition R^TR = I where U be an additional matrix. Now, R can effectively solved by eigen-decomposition of the covariance matrix Σ_x. Then eq. 2 can be redefined as:

$x_{t} = R [\begin{matrix} y_{t} \\ 0 \end{matrix}] + η_{t}$ (3) where 0 means a single column vector and each element be 0. Multiplying R^T on both sides of eq. 3:

$R^{T} x_{t} = R_{t} [\begin{matrix} y_{t} \\ 0 \end{matrix}] + R^{T} η_{t} = [\begin{matrix} y_{t} + R^{T} η_{t} \\ U^{T} η_{t} \end{matrix}]$ (4)

From the eq. 4 it is observed that noise component η_t separated from x_t by rotation matrix R. From the Gaussian properties, η_t = U^Tη_t is a random variable satisfying Gaussian distribution.

Since the rotation matrix R consisting of eigenvectors of Σ_x, we can have

$Σ_{x} R = R Φ$ (5)

Where Φ is a diagonal matrix, the diagonal elements are eigenvalues.

Thus, the covariance matrix of R^Tx_t can be evaluated as:

$\begin{matrix} \frac{1}{s} Σ_{t = 1}^{s} (R^{T} x_{t}) (R^{T} x_{T})^{T} & = \frac{1}{s} Σ_{t = 1}^{s} (R^{T} x_{t} x_{t}^{T} R) \\ = R^{T} Σ_{x} R \\ = [\begin{matrix} λ_{1} & 0 & \dots & 0 \\ 0 & λ_{2} & \dots & 0 \\ ⋮ & ⋮ & ⋱ & ⋮ \\ 0 & 0 & \dots & λ_{r} \end{matrix}] \end{matrix}$ (6)

The eq. 6 indicates that i^th eigenvalue λ_i is the variance obtained on i^th dimension of vector R^Tx_t. It observed that variance on principal dimensions are larger than σ and variances on the redundant dimensions to be equal to σ. Then arranging λ_i in an decreasing order λ₁ ≥ λ₂ ≥ … ≥ λ_r and represent into two vectors S = S₁ ⋃ S₂ where $S_{1} = {λ_{i = 1}^{m}$ is the variance on principal dimensions and $S_{2} = {λ_{i = m + 1}^{r}}$ is variance on redundant dimensions.

Let τ_t[i] denote the i^th element of vector τ_t = U^Te_t, then eigenvalues in the subset S₂ is defined as

$λ_{i} = \frac{1}{s} Σ_{t = 1}^{s} τ_{t} [i]^{2} \forall i \in {m + 1, m + 2, \dots r}$ (7) where τ_t [i]is a random variable following Gaussian distribution N (0, σ²). Since each eigenvalue λ_i ∈ S₂ is itself a random value then median of it is considered as an estimation of the noise variance $σ_{p}^{2}$ in the pixel domain.

2.1.2 Noise estimation in residual domain

The denoised image I_η can be obtained by applying filter H(x,y) is defined as

$I_{η} (x, y) = J (x, y) \otimes H (x, y)$ (8)

The residual image I_r (x, y) consists of the pixel’s noise and obtain by subtracting the denoised image from the original image.

$I_{r} (x, y) = I (x, y) - I η (x, y)$ (9)

Let S_i be the set of irregular objects obtained from the input image, and for each object, the variance of the residual pixels estimated as its noise level as:

$μ (S_{i}) = \frac{1}{| S_{i} |} Σ_{J = 1}^{| S_{i} |} (S_{i} (p_{j}))$ (10) $σ^{2} (S_{i}) = \frac{1}{| S_{i} |} Σ_{J = 1}^{| S_{i} |} (S_{i} (p_{j}) - μ (S_{i}))^{2}$ (11)

μ (S_i) is the pixel average, and $σ_{r}^{2}$ is the variance of the irregular object in residual domain.

Noise variances σ_p in pixel domain & σ_r in the residual domains together used as noise level estimation for localizing tampered region.

2.2 Splicing localization

The work uses Mask-RCNN image segmentation framework [41] to extract individual irregular segments with their bounding boxes. For each segment, the original object is known as foreground, and the remaining bounding box area is called background. Then noise level characteristics are estimated for both foreground and background objects, as discussed in section2.1. Then spliced region is exposed by evaluating cosine similarity among pairs of individual objects.

For any pair of the distinct foreground or background objects, let the noise level estimation be S₁ and S₂. Then the cosine dissimilarity between the objects defined as:

$L_{D} = 1 - \frac{C (S_{1}, S_{2}) + 1.0}{2}$ (12)

where

$C (S_{1}, S_{2}) = \frac{S_{1}^{T} . S_{2}}{∥ S_{1} ∥ . ∥ S_{2} ∥}$ (13)

C (S₁, S₂) is the cosine angle between the two noise levels. The metric L_D gives values in the range [0,1], where the values near to 0 represent similar noise levels and near to 1 represents different noise levels.

Algorithm 1 Algorithm for identifying probable tampered object from Dissimilarity Matrix

Input: Estimated noise levels of N Individual objects of Spliced Image

Output: Tampered Object Find Dissimilarity matrix

1: fori = 1 to Ndo

2: forj = 1 to i - 1 do

3: DM (i, j) = L_D (S_i, S_j)

4: end for

5: end for

Find the pair having maximum dissimilarity

6: for each column in DMdo

7: [COLMAX_j, COLIDX_j] = max (DM_j)

8: end for

9: [cmax, cidx] = max (COLMAX)

10: for each row in DMdo

11: [ROWMAX_i, ROWIDX_i] = max (DM_i)

12: end for

13: [rmax, ridx] = max (ROWMAX)

DM (rmax, cmax) has maximum dissimilarity

Now find which object has maximum dissimilarity

14: RROW = ridx, CCOL = COLIDX (ridx)

15: count = 0 ;

16: for each COLIDXdo

17: if (RROW = COLIDX_j) then

18: count = count + 1

19: end if

20: end for

retrun the tampered object

21: if (count ≥ 0) then

22: T_P = RROW

23: end if

24: returnT_P

After evaluating noise level deviation among pairs of objects, the object which has maximum deviation among all other is estimated by using the proposed localization algorithm 1 and exposed as spliced region.

3 Experimental and performance analysis

In this section, the proposed method is evaluated using two datasets and compare its performance with other existing state-of-art techniques in the literature. Typically, Columbia dataset [39] is a widely used image dataset for localizing image splicing forgery, and it consists of 183 spliced and 180 original raw images without undergoing any preprocessing. The proposed method is tested on spliced images from the Columbia dataset and visually compared with three state-of-art noise level localization methods.

The qualitative evolution of splicing images from the Columbia image dataset are shown in Fig. 2. The first row consists of sample Columbia dataset images and the respective ground truth masks in the second row. The proposed work results present in the third row by highlighting the spliced region with white background. The remaining rows are the results of three state-of-art methods. The fourth row is the results of [18], where the authors used superpixel based noise distribution characteristics to localize tampered regions. The technique localized the spliced object, but there are some false positive and true negative superpixels observed. The fifth row consists of results of [5], where the authors use block-based PCA noise level estimation. The results reveal that lot of blocks are falsely detected as spliced. Since the objects are usually in an irregular shape, it could not cover the entire region. The last row consists of the work of [4], where authors used kurtosis based noise level estimation of each pixel. The technique gives a clue for tampering, but there are many false pixels recognized as spliced pixels. From these, it is evident that the proposed method gives accurate localization results of all images. The state of art methods fails or result in high false alarm rates as the noise differences between original and spliced regions are typically small in the Columbia dataset. The proposed method can distinguish the spliced region even with small noise differences, whereas other methods are not.

Fig. 2

Visual evaluation of proposed method on Colombia data set with [4 , 18].

3.1 Localization accuracy

For evaluating the performance quantitatively, the proposed work uses two metrics: True Positive Rate (TPR) and False Positive Rate (FPR) and are defined as follows:

$TPR = \frac{TP}{TP + FN} * 100$ (14)

$FPR = \frac{FP}{FP + TN} * 100$ (15)

Where TP-True Positive: signifying the number of tampered pixels detected correctly, FP - False Positive: signifying the number of original pixels detected incorrectly, TN - True Negative: signifying the number of original pixels identified correctly, and FN - False Negative: signifying the number of tampered pixels detected incorrectly. TPR gives the rate of correctly recognized pixels as spliced, and FPR is the rate of pixels that are falsely perceived as tampered in the region. The technique is efficient when it expects to have high TPR and low FPR in the results.

Table 1 demonstrates the pixel-level evaluation of the proposed work on the Columbia splicing image dataset. The results show that the proposed method gives accurate localization with less false alarm rate than other existing methods. From Fig. 3, the TPR and FPR of the proposed method show their superiority and stability among all other existing methods.

Table 1

Pixel based accuracy and Comparative results on DSO dataset

	(a)		(b)		(c)		(d)
	TPR	FPR	TPR	FPR	TPR	FPR	TPR	FPR
Proposed Method	84.26	1.4	92.78	0.41	90.90	0.37	87.62	0.29
[18]	89.94	1.2	83.13	0.48	85.87	0.43	82.33	0.32
[5]	78.47	26.26	84.19	1.15	78.53	5.05	82.66	3.25
[4]	40.13	24.90	41.67	28.83	40.27	30.08	48.68	40.92

3.2 Quantitative analysis of Localization accuracy

Most of the noise-based localization methods use datasets consisting of general objects for splicing. In reality, human beings or their faces are used frequently for tampering. To evaluate the performance of the proposed work, DSO spliced imaged dataset [40] is used, consisting of 100 original and 100 spliced high-resolution images of size 1536 x 2048. Fig. 4 consisting of sample images from the DSO dataset and their corresponding masks.

Fig. 3

Qualitative comparison of splicing localization on Columbia dataset.

Fig. 4

sample spliced images from DSO dataset [40] and the corresponding masks.

Table 2 demonstrates the evaluation results of proposed work on the DSO splicing image dataset. The results reveal that the proposed method gives effective localization than other existing methods. From Fig. 5, it is evident that the TPR and FPR of the proposed method is stable.

Table 2

Pixel based accuracy and Comparative results on DSO dataset

	(a)		(b)		(c)		(d)		(e)
	TPR	FPR	TPR	FPR	TPR	FPR	TPR	FPR	TPR	FPR
Proposed Method	100	0.11	91.91	0.36	94.40	0.27	89.12	0.31	91.34	0.23
[18]	89.34	1.91	89.02	0.32	100	5.3	88.94	0.94	92.78	0.31
[5]	79.07	4.13	73.06	4.42	87.53	13.52	80.68	2.84	82.84	3.57
[4]	62.39	16.90	58.63	14.7	56.5	23.04	59.0	16.13	70.68	19.38

Fig. 5

Comparison of splicing localization accuracy on DSO dataset.

For effective localization accuracy, zero-mean white Gaussian noise is added to a particular segment with variance σ² ranging from 0.01 to 0.09 with a step size of 0.02 and spliced it with another original image. In this way, 500 spliced images with different noise levels are generated. Then the proposed method is carried out on those images and compared them with other methods.

Figure 6 are the evaluation results of the sample images. Each row consists of spliced images at different noise levels. The first column (a) represents proposed method results, where each spliced object is visible with high accuracy. Column (b) gives the results of [18], which localizes the spliced object, but as the variance reduces, we find few false superpixels. The results of [5] given in column (c). Since the techniques use block-based segmentation, few blocks falsely identified as spliced, and few others falsely identified as non-spliced. The last column (d) gives the results of [4], where there is a clue for the spliced object, but the false-positive rate is high. The results of Fig. 5 concludes that the proposed method is robust to even small differences in noise levels.

Fig. 6

Visual comparison in noise differences of the proposed method with [4 , 18].

The Comparative results summarized in Fig. 7 of images from the DSO spliced dataset over TPR and FPR concerning different noise levels. Figure (a) representing average TPR with Zero mean white Gaussian noise and (b) average FPR with zero-mean white Gaussian noise. From the results of Fig. 6, it is evident that the proposed method has higher TPR and lower FPR among all other techniques. It also observed that the proposed work’s localization accuracy is efficient and stable over all other existing methods, and no significant degradation detected even if the noise variance is small.

Fig. 7

comparison of splicing localization accuracy on noise differenes from DSO dataset.

3.3 Robustness of proposed method for post-processing operations

The method is useful when it gives stable results even after performing post-processing operations on the spliced image. To evaluate the robustness of the proposed work, Gaussian blur, JPEG compression, and Gamma Correction operations performed on spliced images. The visual evaluation of the proposed method after post-processing operations are shows in Fig. 8. In the first row, a Gaussian Blur filter is applied to make splicing untraceable. From the second row to the fourth row, JPEG compression with compression ratios 90%, 80%, and 70% is applied on spliced images, and in the last row, Gamma Correction (γ = 1) with decreased dynamic range (50 to 200) is applied. The proposed method results are stable and effective in all operations and give very high TPR and very low FPR.

Fig. 8

Visual evaluation of robostness of proposed method.

3.4 Comptational complexity

The method is efficient when the average computational time spent on segmentation, feature computation, and localization is minimal. In this work, after segmenting into individual objects, noise features are estimated from each foreground and background, assuming that the background’s noise level is similar to the whole image. Thereby a lot of computation time is saved. Pairwise dissimilarity is then obtained among objects to localize the tampered region. Table 3 shows the average running time of the proposed work and the other state of the art methods [4 , 18] on Columbia and DSO splicing datasets. Among all, the proposed work takes less time to localize the tampered region.

Table 3
Average Running Time

Columbia Dataset DSO dataset

Proposed 4.6 6.8

method of [18] 17.3 22.8

method of [5] 21.6 27.9

method of [4] 12.6 15.8

	Columbia Dataset	DSO dataset
Proposed	4.6	6.8
method of [18]	17.3	22.8
method of [5]	21.6	27.9
method of [4]	12.6	15.8

3.5 Discussion

The state-of-art methods work on either regular or irregular block-based approaches. To the best of our knowledge, no noise-based techniques use object-based segmentation for localization. The proposed method used object-based segmentation, thereby achieved accurate and efficient localization of the tampered area.

The proposed method is robust to localize even when spliced and original images have small noise differences and achieve higher TPR and lower FPR than the state of the art methods.

The method is also robust in performing post-processing operations such as Gaussian Blur, JPEG compression, and Gamma Correction on a spliced image. However, as the compression ratio decreases further, the method cannot localize accurately, and it considers as future work.

4 Conclusion

This paper proposes an effective method for localizing the spliced image’s tampered region using object-based segmentation. Each segment is then split into the foreground and background objects. Noise level features are estimated for all objected on both pixel and residual domains. Then using cosine similarity among pairs of objects are estimated, and using the proposed algorithm, the tampered region is exposed. The experimental findings demonstrate that the proposed approach is more efficient and reliable than the state-of-art noise-based methods in localizing the manipulated region accurately.

Footnotes

Acknowledgments

The autors expressed deep sense of gratitude to the anonymous reviewers for giving their suggestions to improve the authenticity of the work. The authors also expressed thanks to [5 , ] for providing code and dataset available online.

References

Kobayashi

, Okabe

and Sato

, Detecting forgery from static-scene video based on inconsistency in noise level functions, IEEE Trans Inf Forensics Security5(4) (2009), 883–892.

Farid

, Exposing digital forgeries from JPEG ghosts, IEEE Trans. Inf. Forensics Security4(1) (2009), 154–160.

Mahdian

and Saic

, Using noise inconsistencies for blind image forensics, Image and Vision Computing27(10) (2009), 1497–1503.

Lyu

, Pan

and Zhang

, Exposing region splicing forgeries with blind local noise estimation, International Journal of Computer Vision110(2) (2014), 202–221.

Zeng

, Zhan

and Kang

, Image splicing localization using PCA-based noise level estimation, Multimedia Tools and Applications76(4) (2017), 4783–4799.

Zampoglou

, Papadopoulos

and Kompatsiaris

, Large-scale evaluation of splicing localization algorithms for web images, Multimedia Tools andApplications76(4) (2017), 4801–4834.

Pun

C.M.

and Liu

, Multi-scale noise estimation for image splicing forgery detection, Journal of Visual Communication and Image Representation38 (2016), 195–206.

Chandra Sekhar

P.N.R.L.

and Shankar

T.N.

, Splicing Localization Based on Noise Level Inconsistencies in Residuals of Color Channel Differences, IJRTE8(3) (2019), 764–769.

Chandra Sekhar

P.N.R.L.

and Shankar

T.N.

, Review on Image Splicing Forgery Detection, International Journal of Computer Science and Information Security14(11) (2016), 471–475.

10.

Farid

, Image Forgery Detection A survey, IEEE Signal Processing Magazine26(2) (2009), 16–25.

11.

Pyatykh

, Hesser

and Zheng

, Image noise level estimation by principal component analysis, IEEE Transactions on Image Processing22(2) (2013), 687–699.

12.

Kakar

, Sudha

and Ser

, Exposing digital image forgeries by detecting discrepancies in motion blur, IEEE Transactions on Multimedia13 (2011), 443–452.

13.

Chen

, Zhu

and Heng

P.A.

, An Efficient Statistical Method for Image Noise Level Estimation, IEEE International Conference in Computer Vision (2015), 477–485.

14.

Liu

, Tanaka

and Okutomi

, Single-image noiselevel estimation for blind denoising, IEEE Transactions on Image Processing22(12) (2013), 5226–5237.

15.

Liu

, Pun

C.M.

and Yuan

X.C.

, Digital image forgery detection using JPEG features and local noise discrepancies, The Scientific World Journal (2014), 1–12.

16.

, Lu

, Sun

and Huang

, Digital image splicing detection based on Markov features in DCT and DWT domain, IEEE Transactions on Pattern Recognition45(12) (2012), 4292–4299.

17.

Pan

, Zhang

and Lyu

, Exposing image splicing with inconsistent local noise variances, IEEE international Conference on Computer Photography (2012).

18.

Zhang

, Wang

, Zhang

, et al., Image splicing localization using noise distribution characteristic, Multimed Tools Applications78 (2019), 22223–22247.

19.

, An approach based on object detection for image forensics, 1st International Conference on Industrial Artificial Intelligence (IAI), Shenyang, China, (2019), 1–6.

20.

and Retraint

, An Improved Algorithm for Digital Image Authentication and Forgery Localization Using Demosaicing Artifacts, IEEE Access7 (2019), 125038–125053.

21.

Chen

, Qi

, Wang

, Zheng

, Shim

H.J.

and Shi

, AAn Improved Splicing Localization Method by Fully Convolutional Networks, IEEE Access6 (2019), 69472–69480.

22.

Tripathi

, Kumar

, Tripathi

S.P.

and Yadav

, Automated Image Splicing Detection using Texture based Feature Criterion and Fuzzy Support Vector Machine based Classifier, International Conference on Cutting-edge Technologies in Engineering (ICon-CuTE), Uttar Pradesh, India, (2019), 81–86.

23.

Cozzolino

and Verdoliva

, Noiseprint: A CNN-Based Camera Model Fingerprint, IEEE Transactions on Information Forensics and Security15 (2020), 144–159.

24.

Rao

, Ni

and Zhao

, Deep Learning Local Descriptor for Image Splicing Detection and Localization, IEEE Access8 (2020), 25611–25625.

25.

Phan Ho

A.T.

and Retraint

, Effective images splicing detection based on decision fusion, IEEE International Symposium on Signal Processing and Information Technology (ISSPIT), Ajman, United Arab Emirates, (2020), 1–5.

26.

Selvaraj

and Karuppiah

, Enhanced copy-paste forgery detection in digital images using scale-invariant feature transform, IET Image Processing14(3) (2020), 462–471.

27.

Liu

and Pun

C.-M.

, Locating splicing forgery by adaptive-SVD noise estimation and vicinity noise descriptor, Science Direct-Neuro computing387 (2020), 172–187.

28.

Liu

and Pun

C.-M.

, Exposing splicing forgery in realistic scenes using deep fusion network, Science Direct-Information Sciences526 (2020), 133–150.

29.

Swain

, Very High Capacity Image Steganography Technique Using Quotient Value Differencing and LSB Substitution, Arab J Sci Eng44 (2019), 2995–3004.

30.

Ahilan

, et al., Segmentation by Fractional Order Darwinian Particle Swarm Optimization Based Multilevel Thresholding and Improved Lossless Prediction Based Compression Algorithm for Medical Images, IEEE Access7 (2019), 89570–89580.

31.

Swain

, High Capacity Image Steganography Using Modified LSB Substitution and PVD against Pixel Difference Histogram Analysis, Security and Communication Networks2018 (2018), 995–1004.

32.

Bappy

J.H.

, Simons

, Nataraj

, Manjunath

B.S.

and Roy Chowdhury

A.K.

, Hybrid LSTM and Encoder-Decoder Architecture for Detection of Image Forgeries, IEEE Transactions on Image Processing28(7) (2018), 3286–3300.

33.

Liu

and Pun

C.-M.

, Locating splicing forgery by fully convolutional networks and conditional random field, Signal Processing Image Communication66 (2018), 103–112.

34.

Ali Qureshi

and Deriche

, A review on copy-move image forgery detection techniques, IEEE 11th International Multi-Conference on Systems (2014), 1–5.

35.

Redi

J.A.

, Taktak

and Dugelay

, Digital image forensics: a booklet for beginners, Multimed Tools and Applications51 (2011), 133–162.

36.

Shi

Y.Q.

, Chen

, Xuan

and Su

, Steganalysis Versus Splicing Detection, Digital Watermarking. IWDW 20075041 (2008), 158–172.

37.

Bunk

and Bappy

J.H.

, Detection and Localization of Image Forgeries using Resampling Features and Deep Learning, Computer Vision and Pattern Recognition (2017), 1–9.

38.

, Gkioxari

, Dollar

and Girshick

, Mask R-CNN, EEE International Conference on Computer Vision (ICCV) (2017), 2980–2988.

39.

Hsu

, Chang

, Detecting Image Splicing using Geometry Invariants and Camera Characteristics Consistency, IEEE International Conference on Multimedia and Expo (2006), 549–552.

40.

De Carvalho

T.J.

, Riess

and Angelopoulou

, Exposing digital image forgeries by illumination color classification, IEEE Transactions on Information Forensics and Security8(7) (2013), 1182–1194.

41.

Abdulla

, Mask R-CNN for object detection and instance segmentation on Keras and Tensor Flow, Git Hub Repository, (2017).

Effective splicing localization based on noise level inconsistencies

Abstract

Keywords

1 Introduction

1.1 Related work

1.2 Our contribution

2 Proposed method

2.1.1 Noise estimation in pixel domain

Table 3 Average Running Time Columbia Dataset DSO dataset Proposed 4.6 6.8 method of [18] 17.3 22.8 method of [5] 21.6 27.9 method of [4] 12.6 15.8

4 Conclusion

Footnotes

Acknowledgments

References

Table 3
Average Running Time

Columbia Dataset DSO dataset

Proposed 4.6 6.8

method of [18] 17.3 22.8

method of [5] 21.6 27.9

method of [4] 12.6 15.8