Review: Image Forgery Localization via Fine-Grained Analysis of CFA Artifacts
Emily C. Lennert
image, tampering, forgery, localization, detection, color filter array, CFA, demosaicking, artifacts, resampling
- Ferrara, P.; Bianchi, T.; De Rosa, A.; Piva, A. Image forgery localization via fine-grained analysis of CFA artifacts. IEEE Transactions on Information Forensics and Security. 2012, 7(5), 1566-1577.
The opinions expressed in this review are an interpretation of the research presented in the article. These opinions are those of the summation author and do not necessarily represent the position of the University of Central Florida or of the authors of the original article.
In this study, a method is presented for detecting forgeries in digital images. When a digital image is tampered with, such as if something were added to or removed from the image, the image’s “digital fingerprint” is altered. A digital fingerprint is determined by the unseen statistical properties of an image, which characterize the image acquisition and any subsequent processing.
The method presented relies on traces in the digital fingerprint that are created by the image interpolation, which occurs at two main points. Image interpolation is the estimation of values at a new pixel location using known values at a neighboring pixel location. Interpolation first occurs in image acquisition. As an image is acquired, light enters and is filtered by the color filter array (CFA) before reaching a sensor. The CFA filters the light so that only one color is gathered for each pixel: red, blue, or green. The result is a mosaic-like initial layer, from which missing pixel values for the remaining color layers are interpolated to create a final image. This process is also referred to as demosaicking. Interpolation also occurs when an image is transformed, such as by rotation or scaling, i.e. shrinking or zooming. Pixels within the image are relocated as a result of the transformation, and interpolation is required to assign new values for each color. This process is also referred to as a resampling operation. The interpolation process leaves artifacts, or digital traces, in each pixel.
Generally, an image acquired by a digital camera without additional processing will have demosaicking artifacts at each pixel. Inconsistencies in these demosaicking artifacts, as well as the presence of resampling artifacts, may be indicative of forgery in the inconsistent pixels. This study focuses on the demosaicking artifacts, and how these artifacts may be used to precisely identify a potentially tampered area in the image.
The authors set forth a series of algorithms in the study, which allowed for the detection of interpolation artifacts. This algorithm measures the variance between the acquired pixel, i.e. un-interpolated, and the neighboring interpolated pixel. If the interpolated pixel has not been tampered with, the variance is expected to be high. The variance is high because the acquired pixel does not contain demosaicking artifacts, while the interpolated pixel does. However, in cases of forgery, the processing of the pixel destroys the demosaicking artifacts that were present. In these cases, the variance between the interpolated and initially acquired pixel will be close to zero, since acquired pixels also do not contain demosaicking artifacts. The variance can then be visually represented in a map of the image, showing areas of the image with low variance, i.e. higher likelihood of forgery. Figure 9(c-f) within the study demonstrates the mapping of a forged image based on the proposed algorithm, seen in Figure 9(c), as well as other previously proposed algorithms, seen in Figure 9(d-f). While the algorithm did not determine all instances of forgery in the image, the difference between suspected forged areas and unsuspected areas was distinct. Previously established methods did not provide as clear of distinction between forged and un-forged areas. The authors note that human interpretation of the forgery maps is still required to prevent false positive indications of forgery.
In instances of JPEG image compression, the algorithm was less effective. The ability to localize suspected forgeries decreased considerably when the compressed image quality was below 95%. Below 85% quality, the algorithm was unable to distinguish the presence and absence of demosaicking artifacts.
The algorithm allowed for detection of forgeries at the smallest level, a 2 x 2 block. The authors note that image forgeries smaller than 8 x 8 blocks are uncommon, and it was found that 8 x 8 block resolutions also provided reliable results. The authors concluded that, in instances where the image is uncompressed or of very high compression quality, the presented algorithm provides a valid method for fine point localization of forgeries in digital images.
- An algorithm to detect inconsistencies in demosaicking, i.e. color filter array, artifacts was developed. Inconsistencies, measured by variance, may be indicative of forgery.
- The variance can be visualized on a forgery map. Areas with low variance, i.e. likely forgery, will appear darker, and areas with higher variance, i.e. less likelihood of forgery, will appear lighter.
- The proposed algorithm allows for forgery localization up to a 2 x 2 block resolution, i.e. fine –grained.
- The proposed algorithm performs best on uncompressed JPEG images. Only JPEG images with compression quality values greater than 85%.
In instances of image forgery, it is desirable to locate the exact point of forgery in the image. The proposed algorithm provides a tool that aids in fine-grained identification of likely forged areas in a digital image.
- Using this algorithm, a forgery map may be generated that may allow an analyst to finely locate a forged area within a digital image.