TítuloA Critical Look into Quantization Table Generalization Capabilities of CNN-based Double JPEG Compression Detection
Tipo de publicaciónConference Paper
Year of Publication2022
AutoresRodríguez-Lois, E, Vázquez-Padín, D, Pérez-González, F, Comesaña, P
Conference NameEUSIPCO
Conference LocationBelgrade, Serbia
Palabras claveConvolutional Neural Networks, Double JPEG compression, Image forensics, Source heterogeneity
Resumen Double JPEG compression detection has become a core issue in image forensics, as it provides information about the processing history of the image and its authenticity. Several recent works address this problem by exploiting the potential of CNNs to achieve state-of-the-art performance on test datasets. Unfortunately, those schemes are typically tailored to their specific training conditions and suffer a significant drop of performance in real-life scenarios. This paper aims at assessing the influence of quantization table mismatch (with regards to those seen in training) in the detection of double JPEG compression. Experimental results show inconsistency between different sets of quantization tables, with trained models yielding significantly worse results on unknown sets. This effect is also evident in a more realistic setting, where it appears to be more noticeable for sources falling in operating regions with greater inconsistency.