Digital Image Tamper Detection Based on Multimodal Fusion of Residue Features - Advanced Concepts for Intelligent Vision Systems - page 80

One of the objectives of the work reported here is development of robust and
automatic tamper detection framework for low bandwidth Internet streamed videos
where most of the fingerprints left by tamperer can be perturbed by heavy
compression. However, by fusing multiple image tampering detectors, it could be
possible to uncover the tampering in spite of the heavy compression, as different
detectors use cues and artifacts at different stages of the image formation process. So
if an image lacks certain cues, a complementary detector would be used for making a
decision For example, a copy move forgery might have been created with two source
images of similar quantization settings but very different cameras. In this case, the
copy move forgery can be successfully detected by a different detector. We thus
benefit from having several tamper detection modules at hand rather than only using
the one type of detector. Another advantage of fusing several detector outputs to make
a final decision is that, if one of the detector outputs noisy and erroneous scores, the
other detectors could complement and enhance the reliability of the tamper decision.
Therefore, the advantage of fusion is twofold: to handle images which were subjected
to multiple, diverse types of tampering, and to boost the detection robustness and
accuracy by making different modules work with each other. The challenge, however,
lies in the synergistic fusion of diverse detectors as different detectors are based on
different physical principles and segmentation structures.
We formulate the fusion problem in a Bayesian pattern recognition framework and
use well known Gaussian Mixture Models for the task. The approach is based on
detecting the tamper from the multiple image frames, by extracting noise and
quantization residue features in intra-frame and inter-frame pixel sub blocks (we refer
to pixel sub blocks hence forth in this paper as macro blocks), transforming them into
optimal feature subspace (ICA, CCA or FLD) to extract the maximal correlation
properties, and use GMM classifier to establish possible tampering of video. To
enhance the confidence level of one of the tamper detector, we either perform a fusion
of detector scores (late fusion) or fuse the features first and perform the classification
later (feature fusion). The approach extends the noise residue features reported by Hsu
et al in [11] and expands the pattern recognition formulation proposed by Gopi et al in
[10]. The approach is blind and passive, based on the hypothesis, that typical
tampering attacks such as double compression, re-sampling and retouching can
inevitably disturb the correlation properties of the macro blocks within a frame (intra-
frame) as well as between the frames (inter-frame) and can distinguish the
fingerprints or signatures of genuine video from tampered video frames. The rest of
the paper is organized as follows. Next Section describes the formulation of fusion
problem. The details of the experimental results for the proposed fusion scheme is
described in Section 4. The paper concludes in Section 5 with some conclusions and
plan for further work.
3 Formulating the Fusion Problem
The processing pipeline once the images or video is captured consists of several
stages. First, the camera sensor (CCD) captures the natural light passing through the
optical system. Generally, in consumer digital cameras, every pixel is detected by a
CCD detector, and then passed through different colour filters called Color Filter