Source Camera Identification via Explicit Content–Fingerprint Decoupling with a Dual-Branch Deep Learning Framework

Zijuan Han, Yang Yang, Jiaxuan Lu, Jian Sun, Yunxia Liu, Ngai-Fong Bonnie LawApplied Sciences 2026 · DOI

TL;DR

This paper introduces DCF-Net, a novel dual-branch deep learning framework that explicitly disentangles image content from camera fingerprint features. By leveraging imaging physics and a novel physical consistency constraint, DCF-Net achieves more robust and accurate source camera identification, even under challenging conditions like complex post-processing and similar camera models.

What Problem Does This Paper Address?

In an era where digital images are ubiquitous, verifying their authenticity and origin is crucial, especially in legal and forensic contexts. Source Camera Identification (SCI) aims to determine which device captured an image. Traditional methods relying on EXIF metadata are unreliable as this information is often stripped or altered. Content-based SCI, which extracts unique "fingerprints" left by a camera's sensor and imaging pipeline, is more robust. However, a major challenge is the strong entanglement between image content (what the photo is of) and camera fingerprints (how it was captured). Existing deep learning methods often struggle to suppress content interference, leading to reduced accuracy, especially when images are subjected to various post-processing steps or when cameras share similar hardware. This paper tackles the problem of effectively separating these entangled features to achieve more stable and accurate SCI.

Methods

The core of the proposed method, DCF-Net, is a dual-branch deep learning framework designed to explicitly decouple image content from camera fingerprints.

  • Content Reconstruction Branch: This branch uses an improved U-Net architecture to reconstruct the input image. Its goal is to capture and model scene-related, low-frequency information while suppressing device-specific traces. It includes texture refinement and color correction modules to enhance the quality of the reconstructed content, ensuring it serves as a cleaner baseline for fingerprint extraction.
  • Fingerprint Extraction Branch: This branch utilizes a pre-trained ResNet-50 backbone to directly learn discriminative global features from the original image, focusing on camera hardware characteristics.
  • Noise Residual Analysis Module: This module calculates the difference between the original image and the reconstructed content from the content branch. This residual signal is rich in camera-specific noise patterns and imaging artifacts, making it a valuable source for fingerprint information.
  • Physical Consistency Constraint: A key innovation is the introduction of a loss function that enforces physical consistency. This constraint leverages the multiplicative noise model of image formation (I = Io + Io * K + Θ) to ensure the network's predictions align with the underlying imaging mechanism, guiding the disentanglement process.
  • Loss Functions: The model is trained using a combination of:
  • Classification Loss (Cross-Entropy): To directly learn camera fingerprints.
  • Reconstruction Loss (MSE): To ensure the content branch effectively models image content.
  • Physical Consistency Loss: To enforce the relationship between content, fingerprint, and the original image based on imaging physics.

Datasets

The authors evaluated their method on three publicly available datasets:

  • Dresden: A classic forensic dataset with images from 74 digital cameras across 25 models. For experiments, 18 camera models were used, totaling approximately 15,000 images.
  • Forchheim: A smartphone camera dataset with about 4000 images from 25 devices. This dataset is notable for including post-processed versions of images from five popular social media platforms (Facebook, Instagram, Telegram, WhatsApp, Twitter), making it relevant for real-world scenarios.
  • Vision: A dataset comprising images from 35 smartphone devices (29 models, 11 brands). It features diverse image content and resolutions, with 2900 images selected for experiments.

Key Insights and Ideas

  • Imaging Physics as a Guide: The paper demonstrates that incorporating prior knowledge of how images are formed (the multiplicative noise model) is crucial for effective disentanglement, leading to more stable and interpretable models compared to purely data-driven approaches.
  • Residuals are Key: The difference between the original image and a content-reconstructed image effectively isolates camera-specific noise and artifacts, providing a cleaner signal for fingerprint extraction.
  • Content Matters for Fingerprints: Counter-intuitively, accurately modeling image content (via the content branch) actually helps in isolating camera fingerprints by providing a better baseline to subtract from.
  • Social Media Degradation is a Major Hurdle: While the model performs well on original images, performance drops significantly on heavily compressed social media images, highlighting the challenge of extreme data degradation.
  • Similar Hardware is a Persistent Challenge: Distinguishing between camera models with very similar hardware and imaging pipelines remains a difficult task, even for advanced methods.

Key Results

  • Overall Performance: DCF-Net achieved competitive identification accuracy across all three datasets. For instance, on the Vision dataset, it reached 85.78% accuracy, showing a more constrained performance drop compared to other methods.
  • Ablation Studies:
    • Adding the Noise Residual Analysis module significantly improved performance.
  • Introducing the texture refinement and color correction modules to the content branch led to substantial gains in recognition accuracy.
  • Using all three loss functions (classification, reconstruction, and physical consistency) yielded the best results, demonstrating their complementary benefits.
  • Robustness: The method showed stable performance on the Forchheim dataset, even with post-processed images, achieving an average accuracy of 58.45% across social media platforms. It ranked second among compared methods on this challenging task.
  • Specific Challenges: The model performed well on most camera models in the Dresden dataset, but showed confusion between similar Sony models. On the Vision dataset, it maintained strong performance despite significant degradations.

Limitations and Future Work

  • Severe Degradation: The paper acknowledges that performance degrades significantly on images subjected to extreme compression and post-processing (e.g., Facebook images), where high-frequency fingerprint information is severely attenuated.
  • Similar Camera Models: Distinguishing between camera models with highly similar hardware and imaging processes remains a significant challenge.
  • Future Work: The authors suggest exploring open-set and device-level SCI scenarios to enhance robustness and generalization to unseen cameras. Further research is also needed to address the challenges posed by specific image regions (e.g., over-saturated or highly textured areas) and to improve discrimination among closely matched camera models.

Summary updated

Abstract

In this paper, we propose a source camera identification method based on disentangled feature modeling, aiming to achieve robust extraction of camera fingerprint features under complex imaging and post-processing conditions. To address the severe coupling between image content and camera fingerprint features in existing methods, which makes content interference difficult to suppress, we develop a dual-branch deep learning framework guided by imaging physics. By introducing physical consistency constraints, the proposed framework explicitly separates image content representations from device-related fingerprint features in the feature space, thereby enhancing the stability and robustness of source camera identification. The proposed method adopts two parallel branches: a content modeling branch and a fingerprint feature extraction branch. The content branch is built upon an improved U-Net architecture to reconstruct scene and color information, and further incorporates texture refinement and multi-scale feature fusion to reduce residual content interference in fingerprint modeling. The fingerprint branch employs ResNet-50 as the backbone network to learn discriminative global features associated with the camera imaging pipeline. Based on these branches, fingerprint information dominated by sensor noise is explicitly extracted by computing the residual between the input image and the reconstructed content, and is further encoded through noise analysis and feature fusion for joint camera model classification. Experimental results on multiple public-source camera forensics datasets demonstrate that the proposed method achieves stable and competitive identification performance in same-brand camera discrimination, complex imaging conditions, and post-processing scenarios, validating the effectiveness of the proposed disentangled modeling and physical consistency constraint strategy for source camera identification.
BibTeX
@inproceedings{han2026source,
  author = {Zijuan Han and Yang Yang and Jiaxuan Lu and Jian Sun and Yunxia Liu and Ngai-Fong Bonnie Law},
  title = {{Source Camera Identification via Explicit Content–Fingerprint Decoupling with a Dual-Branch Deep Learning Framework}},
  booktitle = {Applied Sciences},
  year = {2026},
  doi = {10.3390/app16031245},
}
Figure