Digital Image Tamper Detection Based on Multimodal Fusion of Residue Features - Advanced Concepts for Intelligent Vision Systems - page 84

Table 3. (% Accuracy) Performance for noise and quantization residue features and their fusion
for GMM vs. NN classifier

.

% Accuracy GMM

Classifier

SVM
Classifier

NN
Classifier [10[

ICA features ICA features ICA features

Different Residue features and their fusion

83.2 83.5 81.4

f

(Intra-frame noise residue features)

Intra

83.8 83.7 80.6

f

(Inter-frame noise residue features)

Inter

~

77.28 78.28 75.77

f

(Intra-frame quant. residue features)

Intra

~

72.65 73.65 70.53

f

(Inter-frame quant. residue features)

Inter

86.6 85.9 83.22

f

(feature fusion- noise residue)

Intra Inter

~

80.55 81.34 77.23

f

feature fusion- quant residue)

Intra Inter

~

89.56 91.59 83.45

f f

+ (hybrid fusion)

Intra Inter Intra Inter

As can be observed in Table 2 and Table 3, the three classifiers perform differently
for different feature selection techniques. For all three feature selection techniques
GMM and SVM perform much better than the NN classifier. However, for
quantization residue features, the ICA features results in better performance as
compared to CCA features, whereas for noise residue features, CCA gives better
performance. Further, the SVM classifier performs better than the GMM classifier,
for quantization features with ICA feature selection technique. When we perform a
fusion two detectors complement each other and resulting in synergistic fusion with
combination of ICA and SVM processed quantization features and CCA and GMM
processed noise residue features resulting in best performance. The experimental
analysis indicates that for detection of image tampering in low bandwidth video
sequences, we need to use pattern recognition and fusion based formulation, This
formulation allows both linear and nonlinear correlation properties of the tamper
scenarios. In this study we have shown for only two simple types of camera image
post-processing features. However, other features corresponding to interlacing and
de-interlacing artefacts, demosaicing, resampling, touching and blurring need to
examined for characterising the tampering process. This will be the objectives of the
future work.

5 Conclusions

In this paper, we investigated a novel approach for video tamper detection in low-
bandwidth Internet video sequences using a pattern recognition and information
fusion formulation. The approach uses different types of residue features from