Detecting Spliced Images Using Quantum Machine Learning Techniques

Image forgery is a growing concern in today's era of extensive use of social media. It has

changed the way people normally stored, accessed and shared data. Visual imagery has

impacted the routine of documenting evidences and sharing information. Amidst all this

if false information in the form of doctored images are circulated, it misleads people into

drawing false conclusions. Such incidents affect courtroom trials, medical and scientific

research, political campaigns, fashion industry, media and social networking platforms. It

therefore becomes important to differentiate between authentic and forged (or doctored or

tampered) images. The common forms of image forgery include, copy-move and splicing,

along with a series of post-processing operations using sophisticated yet user-friendly image

editing tools for more realistic forgery. Image forensic techniques find traces of image

manipulations by analyzing the image pixels, investigating camera induced and compression

artefacts, studying geometrical and physics based properties of objects captured in the

image. The techniques focus on active and passive detection by classifying authentic vs.

doctored images and extend this to localizing the region on forgery. Earlier work on signal

processing and machine learning techniques have proven effective in detecting the traces of

these image manipulations. Very recently, work in the area of deep learning has showed

remarkable improvements in detecting image manipulations. This work focuses on the

application of (1) machine learning, (2) deep learning and (3) quantum machine learning

techniques to image splicing detection. In (1), features from spliced and authentic images

are engineered by applying the Kekre, discrete cosine, and the hybrid Kekre-discrete cosine

transforms which are then passed onto an assortment of machine learning classifiers to

classify spliced images. For (2), a novel socio-inspired twin convolutional neural network

with a feature-transfer learning approach, named "MissMarple" is proposed to detect traces

of image splicing.