Analysis of Image Quality Characteristics and Processing Artifacts as the Foundation for Deepfake Detection.

Saved in:
Bibliographic Details
Title: Analysis of Image Quality Characteristics and Processing Artifacts as the Foundation for Deepfake Detection.
Alternate Title: Analiza cech jakości obrazu oraz artefaktów przetwarzania jako fundament detekcji deepfake. (Polish)
Authors: Jędrasiak, Karol
Source: Safety & Fire Technology (2657-8808); 2025, Vol. 66 Issue 2, p168-201, 34p
Subject Terms: IMAGE quality analysis, FORENSIC sciences, DIGITAL media, OUTLIER detection
Abstract (English): Aim: The purpose of the article was to empirically verify the hypothesis that image quality descriptors and processing artifacts can provide a stable and interpretable foundation for deepfake detection in real-world distribution conditions (in-the-wild). The study aimed to identify measurable visual characteristics, rooted in the physics of signal acquisition and processing, that allow synthetic content to be distinguished from authentic content with high resistance to platform degradation and recoding manipulation. Project and methodology: The DeepFake RealWorld (DFRW) dataset comprising of 46,371 clips (4,186 authentic and 42,185 synthetic) was developed and utilized, reflecting real-world processing chains and generative models (GAN, diffusion, reenactment, face swap). For each recording, a set of 20 quality descriptors and artifacts were calculated, including BRISQUE, NIQE, PIQE, BLIINDS II, V-BLIINDS, CPBD, Wang-Bovik, PRNU, CFA, and double compression markers. Feature selection was performed without classifiers, by thresholding anomalies defined on the actual class and calculating the p_df, p_real, Δp, and PR indices with FDR control for stability and resistance to platform degradation. Results: Significant differences were found between synthetic and authentic content: on average, p_df = 41.92%, p_real = 26.54%, Δp = 0.15, PR = 1.56. BRISQUE, PIQE, Wang-Bovik, and Laplacian variance, which remained resistant to recoding and mobile filters. PRNU, CFA, and double compression features increased the evidentiary value in high-quality materials. The set of quality characteristics and processing artifacts remained stable under conditions typical for Internet distribution and can be used to calibrate uncertainty and validate forensic systems. Conclusions: The identified quality descriptors and processing artifacts provide an interpretable and robust foundation for deepfake detection, combining perceptual and technical features with acquisition physics. The DFRW dataset enables the construction of hybrid, explainable detectors that combine IQA feature analysis with deep learning models. Future research (DFRWv2) will focus on expanding the dataset to = 500,000 clips with full diffusion model involvement and audio-video multimodality to standardize the reporting of θ, p_df, p_real, Δp, PR, and 95% CI parameters in forensic analyses. [ABSTRACT FROM AUTHOR]
Abstract (Polish): Cel: Celem artykułu była empiryczna weryfikacja hipotezy, że deskryptory jakości obrazu oraz artefaktów przetwarzania mogą stanowić stabilny i interpretowalny fundament detekcji deepfake w warunkach rzeczywistej dystrybucji (ang. in-the-wild). Badanie miało na celu zidentyfikowanie mierzalnych cech wizualnych, zakorzenionych w fizyce akwizycji i przetwarzania sygnału, które pozwalają rozróżniać treści syntetyczne od autentycznych z wysoką odpornością na degradacje platformowe i manipulacje rekodujące. Projekt i metody: Opracowano i wykorzystano zbiór danych DeepFake RealWorld (DFRW) obejmujący 46 371 klipów (4186 autentycznych i 42 185 syntetycznych), odzwierciedlający rzeczywiste łańcuchy przetwarzania i modele generacyjne (GAN, dyfuzja, reenactment, face swap). Dla każdego nagrania obliczono zestaw 20 deskryptorów jakości i artefaktów, w tym BRISQUE, NIQE, PIQE, BLIINDS II, V-BLIINDS, CPBD, Wang-Bovik, PRNU, CFA i markery podwójnej kompresji. Selekcję cech przeprowadzono bez klasyfikatorów, poprzez progowanie anomalii definiowanych na klasie rzeczywistej oraz obliczenie wskaźników p_df, p_real, Δp i PR z kontrolą FDR dla stabilności i odporności na degradacje platformowe. Wyniki: Uzyskano istotne różnice między treściami syntetycznymi i autentycznymi: średnio p_df = 41,92%, p_real = 26,54%, Δp = 0,15, PR = 1,56. Najwyższą skuteczność i stabilność w detekcji deepfake wykazały BRISQUE, PIQE, Wang-Bovik i wariancja Laplasjanu, które pozostawały odporne na rekodowania i filtry mobilne. Cechy PRNU, CFA oraz podwójna kompresja zwiększały wartość dowodową w materiałach wysokiej jakości. Zbiór cech jakości i artefaktów przetwarzania zachował stabilność w warunkach typowych dla dystrybucji internetowej i może być wykorzystany do kalibracji niepewności oraz walidacji systemów forensycznych. Wnioski: Zidentyfikowane deskryptory jakości i artefaktów przetwarzania stanowią interpretowalny i odporny fundament detekcji deepfake, łączący cechy percepcyjne i techniczne z fizyką akwizycji. Zbiór danych DFRW umożliwia budowę hybrydowych, wyjaśnialnych detektorów łączących analizę cech IQA z modelami uczenia głębokiego. Przyszłe badania (DFRWv2) skoncentrują się na rozszerzeniu zbioru do = 500 000 klipów z pełnym udziałem modeli dyfuzyjnych i multimodalnością audio-wideo w celu standaryzacji raportowania parametrów θ, p_df, p_real, Δp, PR i 95% CI w analizach forensycznych. [ABSTRACT FROM AUTHOR]
Copyright of Safety & Fire Technology (2657-8808) is the property of Centrum Naukowo-Badawcze Ochrony Przeciwpozarowej and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database: Complementary Index
Description
Abstract:Aim: The purpose of the article was to empirically verify the hypothesis that image quality descriptors and processing artifacts can provide a stable and interpretable foundation for deepfake detection in real-world distribution conditions (in-the-wild). The study aimed to identify measurable visual characteristics, rooted in the physics of signal acquisition and processing, that allow synthetic content to be distinguished from authentic content with high resistance to platform degradation and recoding manipulation. Project and methodology: The DeepFake RealWorld (DFRW) dataset comprising of 46,371 clips (4,186 authentic and 42,185 synthetic) was developed and utilized, reflecting real-world processing chains and generative models (GAN, diffusion, reenactment, face swap). For each recording, a set of 20 quality descriptors and artifacts were calculated, including BRISQUE, NIQE, PIQE, BLIINDS II, V-BLIINDS, CPBD, Wang-Bovik, PRNU, CFA, and double compression markers. Feature selection was performed without classifiers, by thresholding anomalies defined on the actual class and calculating the p_df, p_real, Δp, and PR indices with FDR control for stability and resistance to platform degradation. Results: Significant differences were found between synthetic and authentic content: on average, p_df = 41.92%, p_real = 26.54%, Δp = 0.15, PR = 1.56. BRISQUE, PIQE, Wang-Bovik, and Laplacian variance, which remained resistant to recoding and mobile filters. PRNU, CFA, and double compression features increased the evidentiary value in high-quality materials. The set of quality characteristics and processing artifacts remained stable under conditions typical for Internet distribution and can be used to calibrate uncertainty and validate forensic systems. Conclusions: The identified quality descriptors and processing artifacts provide an interpretable and robust foundation for deepfake detection, combining perceptual and technical features with acquisition physics. The DFRW dataset enables the construction of hybrid, explainable detectors that combine IQA feature analysis with deep learning models. Future research (DFRWv2) will focus on expanding the dataset to = 500,000 clips with full diffusion model involvement and audio-video multimodality to standardize the reporting of θ, p_df, p_real, Δp, PR, and 95% CI parameters in forensic analyses. [ABSTRACT FROM AUTHOR]
ISSN:26578808
DOI:10.12845/sft.66.2.2025.10