DETECTION OF TAMPERED IMAGES USING AR MODELING AND ARTIFICIAL NEURAL NETWORK

ABSTRACT

Digital image tampering is the process of manipulating photographic images using image processing tools. Digitally forged images are so real that they do not leave any shades of having been tampered with. They can be made uniquely indistinguishable from an authentic image. In this project, an attempt is made to detect tampered images. Digitally Processed Image forgery makes the digital image data highly correlated. This property is exploited by using Autoregressive (AR) coefficients as the feature vector for identifying the location of digital tampering in a sample image. 300 feature vectors from different images were collected and used to train a Back propagation Artificial Neural Network. Percentage of success in identifying the digital forged images is 94.11%.

RESULTS

Image 1

Fig. 1.1 Original image

Tampering: Copy-paste

Fig. 1.2 Tampered block highlighted in white

Image 2

Fig. 2.1 Original image

Tampering: Copy-paste

Fig. 2.2 Tampered blocks highlighted in white (extra flags)

Image 3

Fig. 3.1 Original image with false alarm

Tampering: Scaling

Fig. 3.2 Tampered blocks highlighted in white (with false alarm)

Image 4

Fig. 4.1 Original image with false alarm

Tampering: Copy-paste

Fig. 4.2 Tampered blocks highlighted in white (with false alarm)

Image 5

Fig. 5.1 Original image with false alarm

Tampering: Shading

Fig. 5.2 Tampered block highlighted in white (with false alarm)

Tabulation of Results

Image	Total No. of Blocks = Original + Tampered	Hit = Undetected Original Blocks+ Detected Tampered Blocks	Miss (In Tampered Image)	False Alarm (Original +Tampered)	Percentage Hit = Hit/(Total No. of blocks)x100
1	80(40+40)	80	0	0	100%
2	48	48	0	0	100%
3	48	43	5	1	89.58%
4	198	192	0	6(3+3)	96.96%
5	50	42	0	8(4+4)	84%
				Overall %	94.11%

Conclusion

This method is able to detect tampered sections in digital images in which the tampering done is of the type copy-paste (images 1,2,4) or scaling (image 3) or shading (image 5).

Although there are a few false alarms when some of the original images are tested, the method still detects any tampering done in the same image despite the initial false alarms.