Digital image forensics currently mainly uses PRNU noise as a fingerprint to attribute an image to a particular camera. However PRNU is usually extracted manually using Maximum Likelihood estimation from multiple images from the same source device. In this paper we show that the PRNU can be learned in a data driven fashion using a ResNet based neural network. We also show that it is possible to train a neural network for camera attribution directly on the residual noise, that contains both the PRNU and a random component. We show that both approaches are valid as we obtained results comparable with the state of the art.
cameras, smartphones, and other digital devices often capture photographs and videos that serve as crucial evCamera source identification consists in the attribution idence in criminal investigations. By determining the of an image to the digital camera from which it was camera source, investigators can establish a connection originally captured by using only image features and no between a suspect and a specific crime scene or event. external information. Camera source attribution plays This information becomes vital in establishing a suspect’s a pivotal role in forensic investigations, particularly in presence at a particular location and time, strengthening the realm of digital imagery and video analysis. The the case against them. source information holds significant value for several Camera source attribution also aids in tracking cyber reasons, making it a critical aspect of modern forensic criminal activity, particularly in cases involving child examinations. exploitation, cyberbullying, or online harassment. By
One of the primary reasons camera source attribu- identifying the camera source, law enforcement can trace tion is crucial is its role in authenticating evidence. In the origin of illegal or harmful content, leading to the any criminal investigation, the authenticity of evidence is identification and apprehension of ofenders. This proacparamount. By determining the camera source, the courts tive approach helps protect potential victims and curtail can verify whether an image or video is an original cap- criminal activities. ture or if it has been tampered with or manipulated. This Moreover, standardized camera source attribution verification process is crucial for establishing the chain practices facilitate cooperation among law enforcement of custody and ensuring the evidence presented in court agencies. Criminal activities often transcend jurisdicis reliable and admissible. tional boundaries, and evidence may be collected by dif
Camera source attribution also helps in determining ferent agencies. By following consistent attribution practhe integrity of images. With the advent of sophisticated tices, professionals can seamlessly exchange and analyze photo and video editing software, the risk of forged or visual evidence, enhancing the overall efectiveness of altered visuals has increased. However, each camera criminal investigations. model possesses unique characteristics that act as digital The rest of the paper is structured as follows: section ifngerprints. 2 reviews relevant related works, section 3 describes the
Furthermore, camera source attribution aids investi- dataset, section 4 describes the proposed method, and in gators in linking suspects to crime scenes. Surveillance particular section 4.1 describes a convolutional networks that is trained to classify the source directly from the SYSYEM 2024: 10th Scholar’s Yearly Symposium of Technology, Engi- residual noise, that contains both the PRNU and a ranneering and Mathematics, Rome, December 2-5, 2024 dom component, while section 4.2 introduces a diferent r$afadle.gmryagcuisktr@isp@czd.ipalg(.Run.iGrorymcau1k.i)t; m(Ga.nDd.eMllia@gdisitargis.u);niroma1.it convolutional network that is trained on PRNU isolated. (L. Mandelli); rafal.scherer@pcz.pl (R. Scherer) The results are presented in section 5 and conclusions 0000-0002-3076-4509 (G. D. Magistris); 0009-0008-8447-0830 are drawn in section 6. (L. Mandelli) © 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
Camera attribution techniques are based on the analysis
of the sensor pattern noise (SPN) that is introduced by
the acquisition device. This signal however is the sum of
two components: a random component that depends on
diferent factors in the image acquisition process and a
deterministic component, that depends on intrinsic
properties of the image sensor. This second component, called
Photo Response Non-Uniformity Noise (PRNU), should
be approximately the same in diferent images acquired
by the same device and can be used as a fingerprint of the
device itself. The PRNU component of an image can be
estimated from multiple images coming from the same
device as follows. First the image signal is separated
from the residual noise using a low-pass filter :
= − ()
Then the deterministic component is separated from
the random component by averaging the residual noise
of multiple images [
We tested the proposed method on the Vision dataset [16], that contains labelled images acquired by common devices. In particular the dataset contains flat and natural images for each device, where natural images represent common scenes while flat images represent homogeneus backgrounds, without edges, and can be used to extract the PRNU. Some samples from the dataset are shown in figure 1. We used flat images to compute the residual noise and the PRNU target. We cropped all images, keeping only the top left corner of the image with size 256 × 256.
In this section we describe the proposed method, and in particular we describe the two approaches mentioned in section 1: in section 4.1 the input of the network is the residual noise, with both deterministic and random components, while in section 4.2 the input of the classification network is the PRNU extracted by a second neural network. The two approaches share the the same convolutional neural network for classification described in section 4.3.
( ∈ ) = ((, )) (3) In this section we describe in detail the process we used to isolate the residual noise from the input signal. We used where is the Dirac function and is a thresholding a method based on wavelet decomposition. This method function. is based on the assumption that the wavelet coeficients
In the last few years deep learning has revolution- are modeled as iid Gaussian variables with zero mean
ized the field of computer vision, and in particular many and variance given by a deterministic, unknown spatially
state of the art approaches in computer vision tasks such varying variance field. Given the variance field, the image
as classification, segmentation, etc. are based on con- wavelet coeficients without noise are estimated with a
volutional neural networks[
In this paper we present two approaches: the first con- (, ) runs through an index set J that depends on the sists in a simple convolutional network applied directly decomposition level. 67–72
(6)
ifcient.
ℎ(, ) = ℎ(, )
ˆ2(, ) ˆ2(, ) + 02 and similar for (, ) and (, ), (, ) ∈
Step 4. Repeat Steps 1–3 for each level and each color channel. The denoised image is obtained by applying the inverse wavelet transform to the denoised wavelet coeficients.
For further details we refer the reader to [
but MLE requires multiple images from the same device to extract the PRNU noise from a single image. In this paper we propose a novel approach based on deep learning to extract the PRNU component from the residual noise. According to [21], when the original mapping is closer to an identity mapping, the residual mapping is easier to be optimized. Therefore we propose Resnet-based CNN, and in particular the CSI-CNN architecture[22] shown in Figure 2.
In order for the model to learn the specific target PRNU, the input data is residual noise, the target is PRNU noise extracted with MLE and the loss function is mean squared error(MSE). The PRNU extracted from the Resnet-based CNN is then used to train the Convolutional classifier described in the next section. The complete workflow is illustrated in figure 3.
Our classifier is trained to recognize ten camera classes In this paper we presented two novel deep learning ap(K = 10): Samsung GalaxyS3Mini, Apple iPhone4s, Apple proaches for camera attribution from raw images, obtainiPhone5c, Huawei P9, LG D290, Lenovo P70A, Sony Xpe- ing results comparable with the state of the art. Moreover riaZ1Compact, Microsoft Lumia640LTE, Wiko Ridge4G, we showed that the PRNU, that is the fingerprint of the Xiaomi RedmiNote3, two of these devices are from the image sensor, can be isolated from the input image using same brand. In total, there are 2194 images. As described a data driven approach. We also showed that a neural netin the previous section, we tested two configurations: in work can be trained to classify the target source camera the first the input of the classifier is the residual noise directly from the residual noise, that contains both the while in the second the ResNet based neural network is PRNU and a random component, obtaining even better trained to predict the PRNU from raw images and then results. the obtained PRNU is used to train the classifier. Figure While significant progress has been made in address5 shows the learning curve of the ResNet based neural ing image source attribution, there are still areas that network. Figure 6 shows the confusion matrices for both warrant further exploration and development. A potenconfigurations. With both approaches we reached state tial avenue for future research and improvements of this of the art results, as shown in table 1. I work could be the application of similar techniques to compressed images. This is particularly important because the popular social media platforms compress the uploaded content, hindering the source attribution. Another important line of research could be the application of these models to AI generated content. The advent of AI-generated images indeed, driven by advancements in machine learning and deep learning algorithms, has significant implications across various domains. While these technologies ofer exciting possibilities and creative opportunities, they also raise important ethical, legal, and societal concerns.