Evaluating Deep Learning and Traditional Approaches Used in Source Camera Identification

arXiv:2511.19180 · cs.CV · Submitted 2025-11-24 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Evaluating Deep Learning and Traditional Approaches Used in Source Camera Identification".

Jane: The gist: This paper presents a comparative analysis of three techniques used in source camera identification (SCI): Photo Response Non-Uniformity (PRNU), JPEG compression artifact analysis,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we're starting by looking at the title of this paper, "Evaluating Deep Learning and Traditional Approaches Used in Source Camera Identification," and what that tells us about their overall focus.

Jane: It immediately tells you they aren't just testing one thing; they are comparing three very different ways—PRNU, JPEG artifacts, and CNNs—to see which method is best for identifying the camera source on a smartphone.

Lu: The authors set up this comparison because they acknowledge that as social media filters change images constantly, new techniques like CNNs have emerged, but they want to see how those compare against older sensor-based methods.

Meng: It's important because if a CNN is better at learning features dynamically, it might actually be the right choice for security purposes in the long run.

Lalam: They are comparing PRNU, which comes from sensor noise variations, JPEG artifacts, which come from compression patterns, and CNNs that learn directly from RGB images.

Tom: And they are evaluating each of these three techniques purely on their device classification accuracy against a fixed set of four devices: the Xiaomi Redmi Note 10S, Samsung Galaxy S24, iPhone thirteen Pro Max, and iPhone seventeen Pro.

Jane: That specific setup is crucial because it shows how these different feature extraction methods behave across varying sensor qualities and compression levels.

Lu: It’s a solid foundation for understanding the trade-offs between having an interpretable feature like PRNU versus a model that learns features directly, like the CNN architecture they used.

Meng: I wonder if this fixed set of devices is enough to really prove that a method is robust before we move on to testing it against millions of real-world images.

Lalam: The paper's goal is basically to draw a conclusion about the relative accuracy of these three methods when applied to smartphone images in a controlled way.

The paper's summary: Tom: So, what's the main takeaway from this study on "Evaluating Deep Learning and Traditional Approaches Used in Source Camera Identification"? Basically, they found that while traditional methods still perform well, there is a noticeable gap when you compare them to modern CNN techniques.

Jane: That's right; they see that the JPEG artifact-based classifier gets the highest overall accuracy of zero point nine zero, but then the PRNU-based classifier follows with an accuracy of zero point seven one.

Lu: And interestingly, the CNN-based classifier lags quite a bit behind at only reaching an accuracy of zero point two nine on this specific test set, which is pretty low compared to the others.

Meng: That low number for the CNN suggests that just feeding raw RGB images into a standard network isn't enough to reliably distinguish between these four different camera sensors in this setup.

Lalam: They also noted some specific device behaviors; for instance, the JPEG classifier correctly labeled all test images from the Redmi Note 10S and Samsung Galaxy S24 with a normalized accuracy of one point zero zero on the corresponding diagonal entries.

Tom: So, to put it plainly, traditional feature-based techniques still offer high accuracy numbers compared to the CNN’s lower performance on this specific test set.

Jane: It's about moving toward a system that blends those established methods with deep learning's ability to capture complex visual patterns when we talk about source camera identification.

Lu: They emphasize that combining the strengths of all three techniques would be necessary to achieve a more robust system because no single method is sufficient on its own, especially if we look at real-world data.

Meng: I think we need to see how this hybrid approach actually performs when we move it from the lab bench to actual production traffic, because lab accuracy doesn't always translate well.

Lalam: The paper "Evaluating Deep Learning and Traditional Approaches Used in Source Camera Identification" showed the overall accuracy of zero point nine zero for the JPEG artifact-based classifier as the best result they found across all methods.

The paper's improvements: Tom: So, we’ve seen that traditional methods like JPEG artifacts are actually doing really well when you look at specific devices, but they aren't keeping up with the CNNs in terms of raw accuracy on this test set.

Jane: Right, and now the authors aren't just stopping there; they’re suggesting a few ways to make these different techniques work better together to fix that gap.

Tom: They propose using JPEG artifacts as a pre-filter for the CNN input. So, you use that JPEG feature vector to guide or regularize the first layers of the network, helping it focus on features that aren't messed up by compression.

Lu: That’s an interesting direction because it tries to give the deep learning model some of those interpretable, compression-invariant details from page one when you look at how JPEG works.

Jane: It feels like they want to bridge that performance gap we saw earlier between the traditional and deep learning methods by making the AI learn from both sensor noise and compression patterns at the same time.

Tom: And then there’s this idea for a hybrid system. They suggest concatenating all three feature vectors—the JPEG one, the PRNU pattern vector, and some intermediate features from the CNN—before feeding them into a final classifier.

Meng: Concatenating those vectors sounds like it would leverage both the proven strength of traditional methods and the ability of deep learning to capture complex patterns simultaneously.

Jane: Exactly. You’re not relying on just one fingerprint anymore, but using multiple fingerprints at once for the final decision, which should make things more reliable.

Lu: And they also put forward an adaptive feature selection mechanism. This would mean that we dynamically weight the PRNU and JPEG features based on things like how much compression or noise is actually in the input image.

Tom: That’s smart because it lets you pick "the best overall performance" for any given input instead of just using one fixed vector from page two.

Jane: It suggests moving away from that fixed feature set and toward something more context-aware in how we extract those fingerprints, which is what makes the system flexible.

Lu: I think this whole approach is about building a system where you get the best of both worlds—the interpretability of traditional methods combined with the complexity of deep learning.

Conclusion: Tom: So, to wrap up this paper on "Evaluating Deep Learning and Traditional Approaches Used in Source Camera Identification," we see that the traditional methods like JPEG artifacts still lead the pack in terms of overall accuracy, reaching zero point nine zero.

Jane: Right, and now the authors aren't just stopping there; they’re suggesting a few ways to make these different techniques work better together through hybrid feature fusion.

Tom: They propose using JPEG artifacts as a pre-filter for the CNN input, or creating a hybrid system that concatenates all three feature vectors—the JPEG one, the PRNU pattern vector, and some intermediate features from the CNN.

Lu: And they also put forward an adaptive feature selection mechanism where you dynamically weight those PRNU and JPEG features based on things like compression level or noise in the input image.

Meng: From an engineering standpoint, we need to see if implementing that dynamic weighting actually adds enough accuracy gain to justify the extra computation cost when we ship this to a phone.

Jane: That’s a fair point, Meng; it moves the problem from just achieving a high score on one test set, like their zero point nine zero overall accuracy for "Evaluating Deep Learning and Traditional Approaches Used in Source Camera Identification," to building something that actually runs smoothly in the field.

Tom: Right, so we’ve got these three strong methods being compared and now some solid suggestions for how to combine them into a more powerful system.

Lu: And moving forward, I’m keen to see how these ideas play out when you start testing those hybrid models on much larger datasets than the four devices they used.

Jane: That’s what we'll be looking for next, because the limitations they flag in this paper point directly toward where the real work needs to happen.

Nazarbayev Intellectual School of Science and Mathematics

cs.CV

Submitted: 2025-11-24

Updated: 2026-10-19

Comments: 6 pages, 5 figures, 1 table. Version submitted to SIDe'26 (International Conference on Secure and Intelligent Digital Ecosystems) (Astana, Oct 2026) before peer review. The paper was accepted and will appear in the Springer STEAM-H series; the accepted version uses repeated cross-validation and reports updated results

Code: https://github.com/0x646D/pillow-heif

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 56/100

The gist: The gist: This paper presents a comparative analysis of three techniques used in source camera identification (SCI): Photo Response Non-Uniformity (PRNU), JPEG compression artifact analysis, and

Key concepts

Photo Response Non-Uniformity (PRNU): PRNU is a unique fingerprint derived from the random noise patterns inherent in a camera's sensor. It works by subtracting a smoothed version of an image from the original to isolate these high-frequency, device-specific noise patterns. This pattern acts as a distinct identifier for each specific smartphone camera sensor.
JPEG Artifact Analysis: This technique analyzes the specific distortions introduced during JPEG compression. Since different devices and compression settings leave unique traces in the resulting image blocks, features like block means and variances are extracted from these artifacts. These features help distinguish between images captured by different smartphones.
Convolutional Neural Networks (CNNs): CNNs are deep learning models trained to recognize patterns directly from raw RGB images without manual feature engineering. In this research, the CNN was trained end-to-end to classify images into device classes. However, the model struggled to achieve high accuracy compared to traditional methods, indicating difficulty in learning subtle sensor characteristics.
Source Camera Identification (SCI): SCI is the process of determining which specific camera or device captured a digital image. This is important for verifying image authenticity and tracking content online. The paper explores various mathematical and machine learning techniques used to reliably identify the source device from an image.

Terminology

Accuracy — 1× in this paper · explained in 1 episode
Convolutional Neural Networks (CNNs) — 1× in this paper · explained in 3 episodes

Summary

The gist: This paper presents a comparative analysis of three techniques used in source camera identification (SCI): Photo Response Non-Uniformity (PRNU), JPEG compression artifact analysis, and convolutional neural networks (CNNs) to classify smartphone images.

Introduction and Motivation

Determining the source device of a digital image has become a crucial task in the context of such widespread smartphone usage Smartphone users are able to tamper with images or post them in misleading contexts Hence, source camera identification (SCI) should be used to analyze and track digital images, restoring their authenticity and credibility as sources of information SCI is also important for the development of fields such as multimedia security and intellectual property protection Traditional techniques used in SCI are based on sensor characteristics, specifically in natural images, the dominant part of the pattern noise is the photo-response non-uniformity noise (PRNU) Another class of features used in traditional SCI techniques derives from JPEG compression artifacts As most images circulating online today are shared through social media, which removes metadata and allows users to modify visual features with filters (which distort the patterns used in image investigation), new convolutional neural network (CNN)-based techniques have emerged.

Methods and Data Preparation

The research evaluates three families of source camera identification methods on a dataset consisting of four smartphone devices: Xiaomi Redmi Note 10S, Samsung Galaxy S24, iPhone 13 Pro Max, and iPhone 17 Pro. The data directories are structured as data/redminote10s/, data/samsungs24/, data/iphone13promax/, and data/iphone17pro. All images are loaded using Python and decoded to RGB using Pillow [5] and pillow-heif [6], and converted to NumPy arrays [7]. For the JPEG artifact and PRNU-based pipelines, images are converted to grayscale using the standard ITU-R BT.601 luma transform. For the CNN classifier, the RGB images are resized to 128 × 128 pixels using bilinear interpolation and normalized to the [0, 1] range by dividing by 255. For each method I employ the same stratified 70/30 train-test split (stratified by device class) with a fixed random seed, to make the results directly comparable across the three methods.

Feature Extraction Techniques

The paper details specific feature extraction processes for each method:

(JPEG Artifact-Based Features)

JPEG compression is block-based and operates by applying the two-dimensional discrete cosine transform (DCT) to non-overlapping 8 × 8 blocks, followed by scalar quantization and entropy coding. The feature vector f is formed by concatenating all means and variances of the AC coefficients: f =

"&mu u,v(u,v)̸=(0,0), σ2 u,v(u,v)̸=(0,0)⊤." Before classification, each feature dimension is standardized to zero mean and unit variance.

(PRNU-Based Features)

Photo Response Non-Uniformity (PRNU) can be obtained from a noise residual and used as a device-specific fingerprint. A high-frequency residual is extracted by subtracting a smoothed version of the image: R = I − Gσ1 ∗ I. The PRNU-like pattern is then K(x) = R(x) − Rˆ(x). The pattern K is flattened into a one-dimensional vector g = vec(K), and for computational convenience, a fixed subset of entries is retained as the final feature vector.

(CNN-Based Classifier)

The third approach uses a convolutional neural network trained end-to-end on RGB images to directly learn features for camera identification, in the spirit of [14]. The network architecture involves several layers, including A convolutional layer with 32 filters of size 3 × 3, followed by ReLU activation and 2 × 2 max pooling and an output fully connected layer with 4 neurons (one per device class) and softmax activation. Training uses the categorical cross-entropy loss L = −

- 1/N X N i=1 X 4 k=1 yi,k log pi,k.

Evaluation and Results

The classification performance is measured using the overall accuracy on the test set: Accuracy =

1/Ntest N Xtest i=1 ˆyi = yi. The results show a clear ranking of the methods, where The JPEG artifact-based classifier achieves the highest accuracy of AccJPEG = 0.90, followed by the PRNUbased classifier with AccPRNU = 0.71, while the CNN-based classifier reaches AccCNN = 0.29.

Per-Device Behaviour analysis reveals distinct patterns for each method:

(JPEG Artifact-Based Classifier)

The JPEG classifier correctly labels all test images from redminote10s and samsungs24 with a normalized accuracy of 1.00 on the corresponding diagonal entries. For the two iPhone models, it perfectly recognises iphone13promax (diagonal entry 1.00) and classifies iphone17pro images less correctly, with 0.60 on the diagonal and 0.40 of these images assigned to iphone13promax.

(PRNU-Based Classifier)

The PRNU-based classifier also achieves perfect recognition for the two Android devices: both redminote10s and samsungs24 attain a normalized accuracy of 1.00 on the diagonal. For iphone17pro, 60% of the images are correctly identified, while 40% are assigned to redminote10s.

(CNN Classifier)

The CNN classifier exhibits a distinctly different pattern where All iphone13promax test images are correctly predicted (diagonal entry 1.00), but all iphone17pro images are also assigned to iphone13promax. For the Android devices, the majority of images from both redminote10s and samsungs24 are mapped to the iphone13promax class (0.83 and 0.80 respectively).

Conclusion

The results highlight the superiority of traditional, interpretable feature-based techniques over the CNN-based method. Future research can investigate hybrid approaches by combining the three aforementioned methods. Moreover, a larger dataset can be used to improve the research’s accuracy. The JPEG artifact-based method achieved the highest overall accuracy of 0.90 and exhibited very clear device separation, especially for the two Android smartphones and for the iPhone 13 Pro Max. The PRNU-based method reached 0.71 accuracy and showed strong performance on the Android devices. The CNN method showed an accuracy of 0.29 and a behavior distinct from the other methods. These results highlight the superiority of traditional, interpretable feature-based techniques over the CNN-based method. The JPEG artifact-based classifier shows the highest and most balanced performance across all four devices <ref:2511.

Improvements for AI systems

Incorporate JPEG artifact analysis as a robust pre-filter for CNN input, potentially using its feature vector to guide or regularize early layers of a CNN to focus on compression-invariant features. This could mitigate the performance gap observed between traditional and deep learning methods by ensuring the network learns from both sensor noise (PRNU) and compression patterns simultaneously.
Develop a hybrid SCI system that utilizes the strengths of all three methods by concatenating their feature vectors—the JPEG artifact vector, the PRNU pattern vector, and CNN intermediate feature maps—before feeding them into a final multi-class classifier. This approach aims to leverage the superiority of traditional, interpretable feature-based techniques over the CNN-based method while retaining the deep learning model's ability to capture complex patterns.
Implement an adaptive feature selection mechanism that dynamically weights PRNU and JPEG features based on image characteristics, such as compression level or perceived noise levels. This would allow the system to select the best overall performance for a given input, potentially leading to higher accuracy than relying solely on a fixed feature vector concatenation.
Enhance the CNN architecture by incorporating specialized convolutional layers designed explicitly to detect and suppress PRNU components, as opposed to general feature learning. This addresses the finding that the CNN does not separate all four devices as clearly by forcing it to learn device-specific sensor characteristics rather than just dominant class regularities.

Sources

Adam: A Method for Stochastic Optimization