Towards a Unified Theory of Sensor Pattern Noise: An analysis of dark current, lens effects, and temperature bias in CMOS image sensors

Abstract

Matching images to a discrete camera is of significance in forensic investigation. In the case of digital images, forensic matching is possible through the use of sensor noise present within every image. There exist misconceptions, however, around how this noise reacts under variables such as temperature and the use of different lens systems. This study aims to formulate a revised model of the additive noise for an image sensor to determine if a new method for matching images to sensors could be created which uses fewer resources than the existing methods, and takes into account a wider range of environmental conditions. Specifically, a revised noise model was needed to determine the effects of different lens systems and the impact of temperature on sensor noise. To determine the revised model, an updated literature search was conducted on the background theory relating to CMOS sensors, as the existing work focuses on CCD imaging sensors. This theory was then applied using six off the shelf CMOS imaging sensors with integrated lens systems. An image sensor was examined under scanning electron microscopy and through the use of Energydispersive x-ray spectroscopy the non-uniform structure of individual pixels was visually observed within the sensor. The lens systems were removed and made interchangeable through the use of a 3D printed camera housing. Lens effects were assessed by swapping lens systems between the cameras and using a pinhole lens to remove all optical effects. The temperature was controlled using an Arduino controlled Peltier plate device, and dark current images were obtained at different temperatures using a blackout lens. It was observed that dark current could be used to identify the temperature of the image sensor at the time of acquisition, contrary to the statements in existing literature that sensor pattern noise is temperature invariant. It was shown that the lens system contributes approximately a quarter of the signal power xii used for pattern matching between the image and sensor. Moreover, through the use of targeted signal processing methods and simple ”Goldilocks” filters processing times could be reduced by more than half by sacrificing precision without losing accuracy. This work indicates that sensor pattern noise, while already viable for forensic identification of images to a specific camera, can also be used for identification of an image to a specific lens system and an image sensors temperature. It has also shown that a tool using sensor pattern noise may have a viable future as a forensic method of triage when confronted with large image data sets. Such additional information could prove effective for forensic investigators, intelligence agencies and police when faced with any form of crime involving imaging technology such as fraud, child exploitation or terrorism.