IMAGE SPLICING FORGERY DETECTION AND LOCALIZATION USING FREQUENCY-BASED FEATURES

Transcription

1 Chiew K.T., et al. (Eds.): PGRES 2017, Kuala Lumpur: Eastin Hotel, FCSIT, 2017: pp IMAGE SPLICING FORGERY DETECTION AND LOCALIZATION USING FREQUENCY-BASED FEATURES Thamarai Subramaniam and Hamid Abdullah Jalab University Malaya, Petaling Jaya, MALAYSIA ABSTRACT: Digital images are frequently used in various fields to disseminate information or source of evidence to make certain facts clear and concise. The explosion of digital technologies advancement and the sophistication of image editing software have paved many ways for image forgery. Images are digitally tampered and modified to deceive the receivers of the information. Image splicing is one of the most notorious techniques used to forge images. In image splicing, a new tampered image is created using different fragments from another image(s). Image forgeries are increasingly becoming difficult to detect by human and machine. Thus, it is important to develop an effective and efficient detection method to authenticate the originality of an image. The proposed system will adopt image transform with texture features to extract the features from the image(s) and train the system using three datasets DVMM v1, DVMM v2, and CASIA to detect image splicing forgery. Keywords: Image forensics, forgery detection, image splicing. INTRODUCTION Essential or complex information can be easily conveyed using a single image, as the idiom goes "A picture is worth a thousand words". In our information driven society, digital images are drastically increasing in various fields in order to disseminate information. Truthfulness and integrity of this information that use digital images are sometime very ambiguous. Image tampering may sometime be innocuous but other times it could lead to adverse effects in many fields such as legal system, health care, journalism or social media. Proliferations in current image processing technologies have enabled even casual users to easily manipulate and enhance images with little or no evidence of the alterations. Image alteration or transformation can be effortlessly achieved through various methods or techniques to achieve desired results. Page 33

2 Thamarai Subramaniam and Hamid Abdullah Jalab Image forgeries are achieved through image enhancing, compositing (splicing) and copymove of images. Image retouching (enhancing) used in media industries to manipulate the images to generate a desirable, and more attractive transformation of original images. Though these may not seen as forgery, it s still involved tampering of the original image. Copy-move forgery occurs when a part(s) of an image is copied and pasted to another location in the same original image. Detecting copy-move forgery can be more challenging due to no significant visible changes in texture of the image. Image splicing involves creating a new fake image by combining different image part(s) from one or more images. When manipulation expertly performed spliced regions can be visually imperceptible. Image splicing disrupts higher-order fourier statistics, which can subsequently be used to detect splicing (Farid, 2009). In digital forgeries, the images structure is disturbed resulting inconsistencies between the original and spliced area (Asghar et al., 2016). These inconsistent artifacts are detected during the detections process to verify the originality of the digital image and are mainly divided into active and passive methods. Active authentication of a digital image relies on watermark or digital fingerprint and it requires the knowledge of the original image. In active method, a digital signature or a digital watermark is embedded in the original image that used to authenticate the integrity of the image. Prepared watermark data with a digital key are embedded into original image creating a digital watermark that can be extracted at the receiver side. The image is checked to discover whether digital watermark or digital signature has been compromised (Vyas and Lunagaria, 2014). Passive authentication is used when there is no knowledge of the original image. Tampered image contains modified underlying statistics (Mushtaq and Mir 2014) that may not be visual to human eyes. Passive detection forgery methods divided into 5 main categories; pixel based, format based, camera based, physical environment based and geometry based (Ansari et al., 2014). i. Pixel based techniques detect the statistical irregularities in tampered image pixel are commonly used for tampering. ii. Format Based Techniques deal with image format primarily the JPEG format. JPEG quantization, Double JPEG and JPEG Blocking are used to detect tampering in compressed images. iii. Camera Based Techniques deal with the quantization, colour correlation, gamma correction, white balancing, filtering and JPEG compression that leave an unique signature of type of camera used to capture the images. iv. Physical Based Techniques deal with differences in lighting direction from the environment across the image. Light direction in 2D, light direction 3D and light environment are used to determine the forgery.

3 Image Splicing Forgery Detection And Localization Using Frequency-Based Features v. Geometry Based Techniques deal with measuring the distinct objects in the world and their position relative to the camera. Principal Point and Metric Measurements techniques used to detect forgery using this type of tampering. Generally image forgery detection follows a general framework, which consist the following steps; 1. Image preprocessing: images go through an initial transformation in order to improve classification (Birajdar & Mankar, 2013). Operations such as transforming color image into grayscale, DCT or DWT are used to preprocessing the images. 2. Feature extraction: feature extraction is a method of capturing visual content of images for indexing and retrieval. Feature extracted can be of color, texture or shape features. 3. Feature reduction: features extracted may appear in an inappropriate format or redundant as such feature preprocessing needed to reduce feature dimensionality and in turn improving the computational complexity. 4. Classification: the process of classifying whether an image is authentic and tampered image. 5. Post-processing localization of tampered regions Feature extraction can be done using various algorithms. Feature extraction using spatial based consist of moment, intensity, key points and spatial texture algorithms. Transform based methods consist of frequency, dimensionality reduction and spectral texture. Image transform allows an image to be transform or convert from one domain into another such as frequency. Transform methods enables the features to be easily detected. Hough Transform, Radon Transform, Discrete Cosine Transform (DCT), Discrete Fourier Transform (DFT), and Discrete Wavelet Transform (DWT) are some of the common image transforms used for features extractions. Transform methods further able to reduce feature dimensionality since it use fewer coefficients for overlapping blocks. Image forgeries are increasingly becoming difficult to detect by human and machine. Thus, it is important to develop an efficient detection method to authenticate the originality of an image using frequency-based features. LITERATURE REVIEW Image splicing is a common forgery technique that threatens the integrity and authenticity of digital images. Image splicing involves composition or merging of two or more images to produce a new altered image. Some selected regions undergo a geometric transformation like rotation, scaling, stretchy, skewing or flipping to make tampering imperceptible (Qureshi and Deriche 2015). This could mislead users of thinking or believing something Page 35

4 Thamarai Subramaniam and Hamid Abdullah Jalab that is not the truth. Therefore, the needs to develop an image splicing forgery detection methods are imperative. The detection works by identifying the forged region(s) and localizing the forged area on the tampered images. Numerous researches have been carried out to identify image splicing forgeries. Y. Zhang et al. (2012) proposed to use local binary patterns (LBP) to model magnitude components of 2-D arrays obtained by applying multi-size block discrete cosine transform (MBDCT) to test the images, the resulting LBP features were served as discriminative features for image splicing detection. Their experiments results show a moderate accuracy of % and 88.70% on DVMM Laboratory of Columbia University image dataset. Moreover, the run length based scheme is proposed as well by He Z et al. (2012) to detect splicing. Approximate run length is defined and calculated along the edge gradient direction then SVM is used to classify the authentic and spliced images. The experiments yield reduced computational complexity (extraction time of seconds) but achieved a moderate detection accuracy of 80%. They have concluded that complex textures of spliced images are more likely to be misclassified. Moghaddasi et al., (2014) proposed an improved method of RLRN image splicing detection method by applying a Kernel PCA as dimension reduction % is achieved in the detection accuracy on the DVMM the gray-scale image dataset. Moghaddasi et al.(2014), have extracted SVD-based features and merged with discrete cosine transform (DCT) as a feature extraction method for splicing detection. The results show an accuracy of 78.82% on the DVMM v1 image dataset with 50D feature vector. Liu B., and Pun C., (2015) proposed detecting splicing image using noise discrepancies. Image is segmented using SLIC superpixels algorithm and measured against noise pattern and level and utilized energy-based graph cuts to label spliced area. Their result indicates good detection rate, however images used by them are limited to NIKON D7000 and CANOON 550D noise pattern. Forgery using blur techniques to obscure the tampered area of the image makes detection difficult. Bahrami K. and Kot A. C. (2015) proposed a novel framework for blurred image splicing localization based on the partial blur type inconsistency. Image is partitioned into blocks, the local blur type features are extracted and are classified into out-of-focus or motion blur. A fine splicing localization is applied to increase the precision of regions boundary. The proposed method indicates an accuracy of 94.8%. Zhao et al. (2015) proposed a 2-D non-causal Markov model which captures the underlying dependencies between the current node and its neighbors. Features are extracted using BDCT domain and DMWT domain then 2_D non-causal Markov model is applied and trained

5 Image Splicing Forgery Detection And Localization Using Frequency-Based Features using SVM. The experiments result outperformed some other methods but achieved 93.36% accuracy. Zhang Q., Lu W. and Weng J. (2016) applied Markov model in the block discrete cosine transform (DCT) domain and contourlet transform domain. DCT approach extracts the Markov features of intra-block DCT coefficients, which improved by considering different frequency ranges of each block DCT coefficients. Their experiment results show that their approach can achieve an excellent detection performance on the DVMM dataset with a 96.69% of accuracy. The methods cited above able to detect and localized the splicing image, however there are still some issue need to be addressed. The main problems of the existing methods can be categorized as: 1. System robustness against the post processing operations: Handling of redundancy and higher dimensional of features can effectively increases processing operation thus reducing robustness of detection methods. 2. The accuracy and the false positive rate: Accuracy of detection rate is most imperative in image splicing detection. Low detection, false positive rate and accuracy rate can undermine the effectiveness of the detection methods. 3. Computational cost: Higher features extraction computation time is major drawback is the image splicing detection methods. Therefore, the problems targeted by this research are how would an image splicing detection algorithm is able to efficiently detect tampered images with high detection rate with a reasonable time, and with a low dimensionality features. In order to overcome the challenges that have been identified, there are some research questions need to be considered: i. How to develop new image splicing detection method to detect the spliced images more accurately? ii. How to develop new method with low dimension? iii. How to apply an efficient feature extraction method for image splicing detection? iv. How to test and evaluate the performance of splicing detection approach? Page 37

6 Thamarai Subramaniam and Hamid Abdullah Jalab OBJECTIVE The aim of this research is to develop a low dimensional based image splicing detection method that enhances the accuracy rate with a good enough computational time. The objectives are as follows i. To investigate and analyze the Splicing forgery detection approaches. ii. To design and develop splicing forgery detection method using frequency-based features that is robust to post-processing manipulations iii. To improve the accuracy and reduce false positive of the detection rate iv. To test the proposed methods (transform and texture features) using three standard image datasets: DVMM v1, DVMM v2, and CASIA. v. To evaluate the proposed methods (transform and texture features) using the true positive (TP), true negative (TN), and accuracy (average detection rate). METHODOLOGY A quantitative approach with experimental strategies is used for this research. Previous literatures are located and selected from various resources based on the web of science ranking. Selected resources are critically analyzed and summarized to understand current gap in literatures, definition of system scope, hypothesis and assumptions. Proposed research will be modeled based on the following phases, which encompasses; analysis and preliminary phase, design and prototyping phase, Implementation phase, Validation phase. By applying this methodology, researcher can systematic able to collect necessary data and analysis these data to design relevant prototype and later implement the proposed system. PROPOSED METHODS Researcher proposes the following methods for detecting image splicing forgery. i. Pre-processing: Images require some pre-processing for it to be considered for feature extraction, for an example cropping or converting image format. ii. Feature Extraction: Features extracted from images to construct feature vector for training and classification.

7 Image Splicing Forgery Detection And Localization Using Frequency-Based Features iii. iv. Feature Reduction: Feature vector needs to be constructed with low dimension to reduce training and classification complexity and time. Classification: Images are classified into tampered image or not tampered image base on the classifier. CONCLUSION Advancement in image editing software has made image manipulation a trivial task. Adding, changing and deleting image without any evidence of tampering can be easily carried out. In society where information plays vital role in decision-making, tampered images may lead to poor decision making, planning and controlling by the important player in our society. Image forgeries are increasingly becoming difficult to detect by human and machine, as such it is important to develop robust detection methods that enable to identify the authenticity and credibility of image. Image splicing involves in creating a new image by combining different image part(s) from one or more images. When manipulation expertly performed spliced regions can be visually imperceptible. Image splicing disrupts higherorder Fourier statistics, which can subsequently be used to detect splicing (Farid, 2009). Researcher proposed image transformation methods with texture features to extract the features from image(s) and train the system using three dataset DVMM v1, DVMM v2, and CASIA to detect image splicing forgery. REFERENCES Asghar K., Habib Z., and Hussain M., (2016) Copy-move and splicing image forgery detection and localization techniques: a review. Australian Journal of Forensic Science. Ansari M. D., Ghrera S. P., and Tyagi V., (2014) Pixel-based Image Forgery Detection : A review. IETE Journal Of Education. Vol 55. No Qureshi M. A., and Deriche M., (2015) A bibliography of pixel-based blind image forgery detection techniques. Signal processing: Image Communication Zhang Y., Zhao C., Pi Y., and Li S,. (2012) Revealing Image Splicing Forgery Using Local Binary Patterns of DCT Coefficients. Communication, Signal Processing and System, lecture notes In Electrical Engineering Q. Lianget al.(eds). Springer Science and Business Media New York He Z., Lu W., Sun W,. (2012) Improved Run Length Based Detection of Digital Image Splicing. IWDW 2011, LNCS 7128, Y.Q. Shi, Kim H.J. and Perez-Gonzalez (eds). Springer-Verlag Berlin Heidelberg Page 39

8 Thamarai Subramaniam and Hamid Abdullah Jalab Liu B., and Pun C., (2015) Splicing Forgery Exposure in Digital Image by Detecting Noise Discrepancies. International Journal of Computer and Communication Engineering, Vol 4, No. 1, Bahrami K., and Kot A. C., (2015) Blurred Image Splicing Localization by exposing Blur Type Inconsistency. IEEE Transaction of Information Forensics and Security. Vol 10. No Zhao X., Wand S., Li S., and Li J., (2015) Passive Image Splicing Detection by a 2- D Non-causal Markov Model. IEEE Transactions on Circuits and System for Video Technology. Vol 25 No Nixon M. and Aguado A., (2008) Academic Press. Feature axtraction and image processing. H. Faird (2009) Image forgery Detection A Survey. IEEE Signal Processing Magazine. C. Vyas and M. Lunagaria. A review on methods for image authentication and visual cryptography in digital image watermarking. IEEE International Conference on Computational Intelligence and computing Reseach (ICCIC), India. pp.1-6 Z. Moghaddasi, H. A. Jalab, R. Md Noor, and S. Aghabozorgi, "Improving RLRN image splicing detection with the use of PCA and kernel PCA," The Scientific World Journal, vol. 2014, 2014 Moghaddasi, Z., H.A. Jalab, and R.M. Noor. SVD-based image splicing detection. IEEE International Conference on Information Technology and Multimedia (ICIMU) Malaysia. Fu, D., Shi, Y. Q., and Su, W. (2006). Detection of image splicing based on Hilbert-huang transform and moments of characteristic functions with wavelet decomposition. Digital Watermarking (pp ): Springer Li, C., Ma, Q., Xiao, L., Li, M., and Zhang, A., (2017) Image splicing detection based on Markov features in QDCT Domain. Neuroconputing 228. (pp 29-36): Elsevier Birajdar, G. K., & Mankar, V. H. (2013). Digital image forgery detection using passive techniques: A survey. Digital Investigation, 10(3), doi: