Digital Image Tamper Detection Based on Multimodal Fusion of Residue Features - Advanced Concepts for Intelligent Vision Systems - page 81

Array (CFA). Then, the missing pixels in each color planes are filled in by a CFA
interpolation. Finally, operations such as demosaicing, enhancement and gamma
correction are applied by the camera, and converted to a user-defined format, such as
RAW, TIFF, and JPEG, and stored in the memory.
Since the knowledge about the source and exact processing (details of the camera)
used is not available for application scenarios considered in this work (low-bandwidth
Internet video sequences), and which may not be authentic and already tampered, we
extract a set of residual features for macro blocks within the frame and between adjacent
frames from the video sequences. These residual features try to model and extract the
fingerprints for source level post processing within any camera, such as denoising,
quantization, interlacing, de-interlacing, compression, contrast enhancement, white
balancing, image sharpening etc. In this work, we use only two types of residual
features: noise residue features and quantization residue features.
The noise and quantization residue features were first extracted from 32 x 32 pixel
intra-frame and inter-frame macro blocks of the video sequences. The details of noise
and quantization residue features are described in [3], [4] and [11]. A feature selection
algorithm was used to select those features that exhibit maximal significance. We
used feature selection techniques based on three different techniques: Fisher linear
discriminant analysis (FLD), canonical correlation Analysis (CCA), and Independent
component analysis (ICA). The details of the three feature selection techniques is
described in [12], [13].
4 Experimental Results
The video sequence data base from Internet movie sequences was collected and
partitioned into separate subsets based on different actions and genres. The data
collection protocol used was similar to the one described in [14]. Figure 1 shows
screenshots corresponding to different actions, along with emulation of copy move
tampered scenes and the detection of tampered regions with the proposed approach.
Different sets of experiments were conducted to evaluate the performance of the
proposed feature selection approaches, namely, the ICA, the FLD and the CCA and
their fusion (late fusion or feature fusion) in terms of tamper detection accuracy. The
experiments involved a training phase and a test phase. In the training phase, a
Gaussian Mixture Model for each video sequence from data base was constructed
[15]. In the test phase, copy-move tamper attack was emulated by artificially
tampering the training data. The tamper processing involved copy cut pastes of small
regions in the images and hard to view affine artefacts. Two different types of tampers
were examined. An intra-frame tamper, where the tampering occurs in some of the
macro blocks within the same frame, and inter-frame tamper, where macro blocks
from adjacent frames were used. However, in this paper, we present and discuss
results for the intra-frame tamper scenario only.
As can be seen from Table 1, which show the tamper detection results in terms of
% accuracy, the performance of noise residue and quantization residue features
without feature selection, the improvement achieved by using feature selection
techniques, and the robustness achieved by fusing the sub space features (feature level
fusion) or the scores. We compared the performance of proposed feature selection and
fusion techniques with feature selection based on autoregressive coefficients and
neural network classification proposed by Gopi et al in [10].