In forensic investigation of criminal cases like child pornography, image forgery, identity theft, steganography, movie piracy, insurance claims, and other cases of scientific frauds, some of the most significant challenges may be to detect the origin of an image or the photographing camera, detect forged images or hidden messages in images from retrieved digital evidence. There has been much interest in developing camera fingerprints for the forensic task of digital camera identification; that is, to be able to tie an image to it's photographing camera with high certainty or good assurance metrics, specially when the camera is not present in the crime scene. Inspired by the existing approaches of camera fingerprint forensics, this paper explores a novel approach for camera identification, based on PRNU noise fingerprint, using Artificial Neural Network (ANN) algorithm. While statistical algorithms produce probabilistic inferences based on statistical problem data, artificial neural network algorithm learns features about the problem from training data. Based on correctness of feature representations and complex mathematical processing on the training data, the neural network is able to learn or approximate any non-linear distribution very easily. As it trains on different examples, it's generalization performance on new inputs improves. In currently proposed work, first the reference fingerprint and test fingerprint are estimated based on a simple kernel based processing algorithm for PRNU coefficient estimation. Then an artificial neural network is set up in C programming language for PRNU pattern recognition based on the estimated feature values from the reference pattern data. The network is presented with training inputs and desired outputs, and based on formulated assumptions and hypothesis described in later sections, the expectation is that the ANN will be able to recognize PRNU fingerprint in images taken by the same camera whose fingerprint the ANN got trained on. A low Mean Square Error (MSE) during ANN training and testing is an indication that the ANN could report with high confidence, a match between the camera fingerprint pattern and the pattern in test image. Multilayer Perceptron (MLP) ANN with single hidden layer is proved to be a universal non-linear function approximator and can be applied to solve any complex non-linear problem. Current approach uses back propagation MLP ANN algorithm for fingerprint detection or camera identification.

Library of Congress Subject Headings

Pattern recognition systems; Digital cameras--Identification; Machine learning; Neural networks (Computer science); Computer algorithms

Publication Date


Document Type


Student Type


Degree Name

Computing Security (MS)


Yin Pan

Advisor/Committee Member

Sumita Mishra

Advisor/Committee Member

Roger Dube


Physical copy available from RIT's Wallace Library at TK7882.P3 B46 2013


RIT – Main Campus