Denoising Diffusion Probabilistic Models for Source Camera Identification in Image Forensics

Zahra Farzadpour, Farah Nafees Ahmed, Fouad Khelifi
Proceedings of the Fourth UK AI Conference 2026, PMLR 348:68-77, 2026.

Abstract

Source Camera Identification (SCI) is a forensic task to determine the physical imaging device for a given digital image or video. The problem has raised concern with the proliferation of digital cameras in consumer devices, the ease with image capturing, editing, and sharing at scale across social media and messaging platforms, and the operational need to attribute illicit content in important evidence cases like child abuse to specific recording devices in criminal investigations. The main signal for SCI is Photo-Response Non-Uniformity (PRNU), classically extracted via a denoising filter such as the Wiener filter, which degrades substantially on small image patches. We propose a combined framework in which a per-camera Denoising Diffusion Probabilistic Model (DDPM) acts as a camera-dependent residual extractor, replacing the classical filter, and Linear Discriminant Analysis (LDA) exploits the full NCC score vector across all candidate cameras for the final identification decision. Evaluated on three benchmarks at $128\times128$ patches, our method achieves macro-averaged balanced accuracies of 93.74%, 93.84%, and 92.50% on the Northumbria, Dresden, and VISION datasets respectively, outperforming the Wiener filter on all three benchmarks.

Cite this Paper

BibTeX
@InProceedings{pmlr-v348-farzadpour26a, title = {Denoising Diffusion Probabilistic Models for Source Camera Identification in Image Forensics}, author = {Farzadpour, Zahra and Ahmed, Farah Nafees and Khelifi, Fouad}, booktitle = {Proceedings of the Fourth UK AI Conference 2026}, pages = {68--77}, year = {2026}, editor = {Benford, Alistair and Büyükateş, Baturalp and Cabrera, Christian and Kiden, Sarah and Salili-James, Arianna and Zakka, Vincent and Zhou, Feng}, volume = {348}, series = {Proceedings of Machine Learning Research}, month = {29--30 Sep}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v348/main/assets/farzadpour26a/farzadpour26a.pdf}, url = {https://proceedings.mlr.press/v348/farzadpour26a.html}, abstract = {Source Camera Identification (SCI) is a forensic task to determine the physical imaging device for a given digital image or video. The problem has raised concern with the proliferation of digital cameras in consumer devices, the ease with image capturing, editing, and sharing at scale across social media and messaging platforms, and the operational need to attribute illicit content in important evidence cases like child abuse to specific recording devices in criminal investigations. The main signal for SCI is Photo-Response Non-Uniformity (PRNU), classically extracted via a denoising filter such as the Wiener filter, which degrades substantially on small image patches. We propose a combined framework in which a per-camera Denoising Diffusion Probabilistic Model (DDPM) acts as a camera-dependent residual extractor, replacing the classical filter, and Linear Discriminant Analysis (LDA) exploits the full NCC score vector across all candidate cameras for the final identification decision. Evaluated on three benchmarks at $128\times128$ patches, our method achieves macro-averaged balanced accuracies of 93.74%, 93.84%, and 92.50% on the Northumbria, Dresden, and VISION datasets respectively, outperforming the Wiener filter on all three benchmarks.} }
Endnote
%0 Conference Paper %T Denoising Diffusion Probabilistic Models for Source Camera Identification in Image Forensics %A Zahra Farzadpour %A Farah Nafees Ahmed %A Fouad Khelifi %B Proceedings of the Fourth UK AI Conference 2026 %C Proceedings of Machine Learning Research %D 2026 %E Alistair Benford %E Baturalp Büyükateş %E Christian Cabrera %E Sarah Kiden %E Arianna Salili-James %E Vincent Zakka %E Feng Zhou %F pmlr-v348-farzadpour26a %I PMLR %P 68--77 %U https://proceedings.mlr.press/v348/farzadpour26a.html %V 348 %X Source Camera Identification (SCI) is a forensic task to determine the physical imaging device for a given digital image or video. The problem has raised concern with the proliferation of digital cameras in consumer devices, the ease with image capturing, editing, and sharing at scale across social media and messaging platforms, and the operational need to attribute illicit content in important evidence cases like child abuse to specific recording devices in criminal investigations. The main signal for SCI is Photo-Response Non-Uniformity (PRNU), classically extracted via a denoising filter such as the Wiener filter, which degrades substantially on small image patches. We propose a combined framework in which a per-camera Denoising Diffusion Probabilistic Model (DDPM) acts as a camera-dependent residual extractor, replacing the classical filter, and Linear Discriminant Analysis (LDA) exploits the full NCC score vector across all candidate cameras for the final identification decision. Evaluated on three benchmarks at $128\times128$ patches, our method achieves macro-averaged balanced accuracies of 93.74%, 93.84%, and 92.50% on the Northumbria, Dresden, and VISION datasets respectively, outperforming the Wiener filter on all three benchmarks.
APA
Farzadpour, Z., Ahmed, F.N. & Khelifi, F.. (2026). Denoising Diffusion Probabilistic Models for Source Camera Identification in Image Forensics. Proceedings of the Fourth UK AI Conference 2026, in Proceedings of Machine Learning Research 348:68-77 Available from https://proceedings.mlr.press/v348/farzadpour26a.html.

Related Material