When CNNs Truly Outperform PCE on Source Camera Attribution?
IEEE Access · 2026 · 14 · 67465-67475
DOI: 10.1109/access.2026.3689977 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Andrea Montibeller und 3 weiteren Autor:innen. Die Arbeit nennt 26 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: FEDER and Xunta de Galicia through the “Grupos de Referencia Competitiva”, FEDER and MCIN/AEI through FELDSPAR, “NextGenerationEU/PRTR” through TRUFFLES.
Abstract (Originalsprache)
Source camera attribution is a central problem in multimedia forensics, traditionally addressed using the Peak-to-Correlation Energy (PCE) statistic. A recent study by Mandelli et al. claimed that trainable CNN-based detectors substantially outperform the PCE. In this work, we show that this conclusion is affected by a methodological issue: the evaluation setup adopted for the PCE places it at an inherent disadvantage, leading to an unfair performance comparison. After clarifying the correct use of the PCE detector and re-establishing a fair baseline, we introduce two improvements—a weighted cross-entropy loss to mitigate class imbalance, and a finetuning procedure that better reflects real-world scenarios—to enhance the CNN-based approaches. With these enhancements, some of the architectures proposed by Mandelli et al. indeed surpass a properly implemented PCE across a broader range of operating conditions. The gains are particularly pronounced for medium and large test image crop sizes.
Förderung
- FEDER and Xunta de Galicia through the “Grupos de Referencia Competitiva”
- FEDER and MCIN/AEI through FELDSPAR
- “NextGenerationEU/PRTR” through TRUFFLES
- FF4ALL and SERICS through the MUR National Recovery and Resilience Plan funded by European Union-NextGenerationEU
Zitieren
APA
Montibeller, A., Masciopinto, M., Pérez-González, F., & Boato, G. (2026). When CNNs Truly Outperform PCE on Source Camera Attribution?. IEEE Access, 14, 67465-67475. https://doi.org/10.1109/access.2026.3689977
Vancouver
Montibeller A, Masciopinto M, Pérez-González F, Boato G. When CNNs Truly Outperform PCE on Source Camera Attribution?. IEEE Access. 2026;14:67465-67475. doi:10.1109/access.2026.3689977
BibTeX
@article{Montibeller2026,
author = {Montibeller, Andrea and Masciopinto, Miguel and Pérez-González, Fernando and Boato, Giulia},
title = {When CNNs Truly Outperform PCE on Source Camera Attribution?},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {67465--67475},
doi = {10.1109/access.2026.3689977}
}
EUVIMED