Source Camera Identification Using Large Components of Sensor Pattern Noise

Abstract

Digital image forensics has attracted a lot of attention recently for its role in identifying the origin of digital image. Although different forensic approaches have been proposed, one of the most popular approaches is to rely on the imaging sensor pattern noise, where each sensor pattern noise uniquely corresponds to an imaging device and serves as the intrinsic fingerprint. The correlation-based detection is heavily dependent upon the accuracy of the extracted pattern noise. In this work, we discuss the way to extract the pattern noise, in particular, explore the way to make better use of the pattern noise. Unlike current methods that directly compare the whole pattern noise signal with the reference one, we propose to only compare the large components of these two signals. Our detector can better identify the images taken by different cameras. In the meantime, it needs less computational complexity.

On the SPN Estimation in Image Forensics: A Systematic Empirical Evaluation

“…The authors in [41] proposed to only use the large components in the estimated PRNU to cut down the overall random noise. In theory, the large components carry more of the signal of interest in comparison to small components that are mainly random noise.…”

Sensor fingerprint digests for fast camera identification from geometrically distorted images

“…‡ The idea of speeding up the camera identification using sensor fingerprints by restraining the computations to the largest components in the fingerprint was independently proposed by Hu et al 11 and by Goljan et al.…”

Enhancing Sensor Pattern Noise via Filtering Distortion Removal

“…3) Correlation: Transform the wavelet coefficients back onto the spatial domain and calculate the correlation with the reference pattern noise according to Equation (3). Notice that the reference pattern noise is first represented in the wavelet domain and only the coefficients at the selected locations are preserved.…”

“…proposed a maximum likelihood estimation (MLE) of the reference SPN from several residual images. Hu et al [3] argued that the large or principal components of noise residue are more robust against random noise, so instead of using the full-length SPN, only a small portion of the largest components are involved in the calculation of correlation. Kang et al [4] introduced a camera reference phase SPN to remove the periodic noise and other non-white noise contamination in the camera fingerprint.…”

“…It is widely accepted that the choice of denoising filters is very important [3,8,9] and the quality of SPN can be improved by choosing better denoising filters. However, in their pursuit for better SPN, researchers usually ignore the fact that even the "best" denosing filter can inflict distortion on SPNs.…”