Abstract
Photo Response Non-Uniformity (PRNU) serves as a unique "digital fingerprint" for imaging sensors, enabling forensic experts to link images to specific devices. This paper explores the security implications of PRNU in iris biometrics, demonstrating a targeted attack that spoofs a sensor's identity without compromising the biometric recognizability of the iris. We propose an iterative, patch-based perturbation engine guided by a candidate image from the target sensor. Our results across 11 diverse datasets show a spoof success rate of 89.21% while maintaining iris matching performance within a 1% deviation from the original.
Introduction
In digital forensics, every camera sensor leaves an invisible, unique noise pattern on images known as Photo Response Non-Uniformity (PRNU). Caused by microscopic manufacturing variations in silicon, PRNU has become a gold standard for device authentication. In iris recognition systems, this provides a layer of two-factor authentication: verifying both the individual's identity and the legitimacy of the capturing device.
However, the reliability of this forensic tool depends on the assumption that PRNU patterns cannot be easily forged. This research investigates whether an adversary can "swap" the sensor fingerprint of an iris image to match a target device while ensuring the iris remains identifiable by automated matching systems.
The Challenge of PRNU Spoofing
The primary difficulty in spoofing PRNU is the "balancing act" between forensic deception and biometric integrity. Injecting too much noise to fool a sensor classifier risks distorting the intricate textures of the iris, leading to a failure in the biometric matcher. Conversely, insufficient perturbation fails to achieve a successful spoof.
Proposed Methodology
The authors developed a two-stage targeted attack framework designed to navigate this trade-off:
- Candidate Selection: The algorithm begins by selecting a "candidate image" from a gallery of the target sensor's images. By choosing a candidate with a similar intensity distribution to the source image, the injected noise appears more natural and aligns better with the target sensor's characteristics.
- Iterative Perturbation Engine: Rather than applying a global filter, the method uses a localized, iterative approach. The image is divided into a grid, and $10 \times 10$ pixel patches are selected at random. For each patch, the algorithm calculates two potential modifications (positive and negative) based on the candidate image's features. It adopts the change that maximizes correlation with the target sensor's PRNU.
This process continues until the image's correlation with the target sensor reaches a predefined threshold. This patch-by-patch refinement ensures that the high-frequency iris textures required for recognition are preserved.
Experimental Results
The study utilized 11 standard iris datasets (including BioCOP, IITD, and CASIA) and focused on 20 source-to-target sensor pairs involving hardware from Aoptix, JPC, IC, Cogent, and LG.
Spoofing Effectiveness
The proposed method achieved a Spoof Success Rate (SSR) of 89.21%, significantly outperforming traditional baselines such as PRNU Injection (57.60%) and PRNU Substitution (32.75%). This high success rate indicates that the iterative approach is highly effective at misleading sensor identification systems.
Biometric Preservation
Crucially, the attack does not degrade the utility of the iris image for identification. Using the VeriEye iris matcher, the researchers generated Receiver Operating Characteristic (ROC) curves comparing original and spoofed images.
As shown in the ROC curves, the performance of spoofed images remains almost indistinguishable from the originals, typically staying within a 1% performance margin.
Conclusion
This research demonstrates a significant vulnerability in PRNU-based forensic systems. By using a guided, iterative optimization, an attacker can effectively forge a sensor's digital fingerprint without losing the biometric identity of the subject. These findings suggest that forensic experts must look beyond simple PRNU correlation and develop more robust detection methods capable of identifying the subtle artifacts left by patch-based perturbations.
Summary updated