A Deep Learning Framework for Image Authentication: An Automatic Source Camera Identification Deep-Net

• Published in: Arabian journal for science and engineering (2011) 2023-02, Vol.48 (2), p.1207-1219
• Main Authors: Bharathiraja, S., Rajesh Kanna, B., Hariharan, M.
• Format: Article
• Language: English
• Subjects: Authentication; Cameras; Deep learning; Digital imaging; Engineering; Feature extraction; Forensic sciences; Humanities and Social Sciences; multidisciplinary; Nonuniformity; Research Article-Computer Engineering and Computer Science; Science
• ISSN: 2193-567X; 1319-8025; 2191-4281
• DOI: 10.1007/s13369-022-06743-3

The main aim of digital image forensics is to validate the authenticity of images by identifying the camera that captured the image and finding traces of any alteration in the spatial content. The majority of the existing literature focuses on manual extraction of intrinsic camera features such as lens aberration, sensor imperfections, pixel non-uniformity, color filter array type, and so on. These handcrafted features are analyzed and characterized as a unique signature for detecting the camera and authenticating the image which is recorded from the device. To facilitate an end-to-end automated forensics analysis, the current research explores the ability of a novel deep learning framework to learn the intrinsic signature of a selected camera model. The proposed deep convolutional network performs source camera identification (SCI). It contains two functional blocks, namely, esidual noise feature extractor (RNFE) and Feature Modulator (FM). To extract the noise pattern from camera images, the RNFE module analyses the debayered image using a U-Net. The generated noise residue is then modulated through a CNN pipeline to an embedding vector. Triplet loss function is used to train the proposed SCI network such that, the images captured from the source cameras are located closer to each other than images from different cameras. Experimental results demonstrate that the CNN achieves a 97.59% F-score and 97.01% recall, on par with state-of-the-art. Hence, the unified architectural representation of the proposed deep-net could be treated as a generic deep net framework in learning the sensor pattern noise (SPN) fingerprint of a camera model.