Abstarct: With the advancement of digital technology, digital media can be recorded from the cameras of various devices and easily shared in cyberspace and social networks. In this context, video and video manipulations have become very common and increasingly dangerous for individuals and society as a whole. Identifying the camera model is the first step in assessing the authenticity of images in forensic research on digital images. Source camera identification is the topic of this dissertation. In most related articles, various deep learning architectures were designed to extract the noise pattern as the fingerprint of each camera. Each image contains noises that are unique to each camera baxsed on the design of the lens system. This noise pattern will not be lost due to image editing and manipulation, and the ultimate goal of extracting the noise pattern is to identify the camera model of each image that is used in the discussion of image criminology. Specifically, this dissertation analyzes CNN data processing and architecture. Data preprocessing does not always help identify the source camera, as some data may be lost during the preprocessing process and the efficiency of the source camera identification system may be reduced. Therefore, choosing the right preprocessors is important. The proposed method of this dissertation focuses on data processing using the functions of increasing the image resolution, increasing the contrast in the image and recognizing the edge in the image. Next, the determination of deep neural network structure for feature extraction was examined. A deep neural network structure consisting of two convolutional laxyers, a maximal retractor laxyer and two fully connected bilxayers was used. The softmax function and the SVM support vector machine at the end of deep neural network structures were used to classify the datasets. In this dissertation, to prevent overfitting, the DropConnect method was used instead of the DropOut method, in order to extract a more accurate noise pattern by preserving more neurons than the DropOut method, and to improve the detection efficiency of the source camera. At the model and sensor levels, the deep neural network structure with softmax function was able to correctly identify 98.5% and 95.4% of the test data, respectively. Also, the structure of deep neural network with support vector machine correctly identified 98.3% and 93.2% of the test data, respectively. In this way, the proposed methods have been more accurate than other sources.