Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Analyzing Fairness in Deepfake Detection With Massively Annotated Databases

Ying Xu, Philipp Terhörst, Kiran Raja, Marius Pedersen

TL;DR

This work tackles the fairness of Deepfake detectors by highlighting how biased training data propagate to detection outcomes. It introduces a large-scale annotation pipeline that transfers 47 attributes to five popular Deepfake datasets, yielding about $65.3 ext{M}$ labels, and it analyzes bias across three backbone models on four datasets using 31 attributes. The authors propose a corrected relative performance metric $CRP(a)$ to separate testing-data imbalance from attribute-driven bias, enabling robust bias quantification. Their findings reveal strong attribute-driven biases and limited diversity in the datasets, underscoring the need for balanced data and bias-aware detectors to improve generalizability and security across real-world populations. Public release of the annotated datasets and code provides a valuable resource for future bias mitigation and fair Deepfake-detection research.

Abstract

In recent years, image and video manipulations with Deepfake have become a severe concern for security and society. Many detection models and datasets have been proposed to detect Deepfake data reliably. However, there is an increased concern that these models and training databases might be biased and, thus, cause Deepfake detectors to fail. In this work, we investigate factors causing biased detection in public Deepfake datasets by (a) creating large-scale demographic and non-demographic attribute annotations with 47 different attributes for five popular Deepfake datasets and (b) comprehensively analysing attributes resulting in AI-bias of three state-of-the-art Deepfake detection backbone models on these datasets. The analysis shows how various attributes influence a large variety of distinctive attributes (from over 65M labels) on the detection performance which includes demographic (age, gender, ethnicity) and non-demographic (hair, skin, accessories, etc.) attributes. The results examined datasets show limited diversity and, more importantly, show that the utilised Deepfake detection backbone models are strongly affected by investigated attributes making them not fair across attributes. The Deepfake detection backbone methods trained on such imbalanced/biased datasets result in incorrect detection results leading to generalisability, fairness, and security issues. Our findings and annotated datasets will guide future research to evaluate and mitigate bias in Deepfake detection techniques. The annotated datasets and the corresponding code are publicly available.

Analyzing Fairness in Deepfake Detection With Massively Annotated Databases

TL;DR

This work tackles the fairness of Deepfake detectors by highlighting how biased training data propagate to detection outcomes. It introduces a large-scale annotation pipeline that transfers 47 attributes to five popular Deepfake datasets, yielding about labels, and it analyzes bias across three backbone models on four datasets using 31 attributes. The authors propose a corrected relative performance metric to separate testing-data imbalance from attribute-driven bias, enabling robust bias quantification. Their findings reveal strong attribute-driven biases and limited diversity in the datasets, underscoring the need for balanced data and bias-aware detectors to improve generalizability and security across real-world populations. Public release of the annotated datasets and code provides a valuable resource for future bias mitigation and fair Deepfake-detection research.

Abstract

In recent years, image and video manipulations with Deepfake have become a severe concern for security and society. Many detection models and datasets have been proposed to detect Deepfake data reliably. However, there is an increased concern that these models and training databases might be biased and, thus, cause Deepfake detectors to fail. In this work, we investigate factors causing biased detection in public Deepfake datasets by (a) creating large-scale demographic and non-demographic attribute annotations with 47 different attributes for five popular Deepfake datasets and (b) comprehensively analysing attributes resulting in AI-bias of three state-of-the-art Deepfake detection backbone models on these datasets. The analysis shows how various attributes influence a large variety of distinctive attributes (from over 65M labels) on the detection performance which includes demographic (age, gender, ethnicity) and non-demographic (hair, skin, accessories, etc.) attributes. The results examined datasets show limited diversity and, more importantly, show that the utilised Deepfake detection backbone models are strongly affected by investigated attributes making them not fair across attributes. The Deepfake detection backbone methods trained on such imbalanced/biased datasets result in incorrect detection results leading to generalisability, fairness, and security issues. Our findings and annotated datasets will guide future research to evaluate and mitigate bias in Deepfake detection techniques. The annotated datasets and the corresponding code are publicly available.
Paper Structure (25 sections, 3 equations, 5 figures, 19 tables)

This paper contains 25 sections, 3 equations, 5 figures, 19 tables.

Table of Contents

  1. Introduction
  2. Societal and Technological Aspects
  3. Contributions
  4. Related Work
  5. Deepfake Detection
  6. Deepfake Datasets
  7. Analysing Bias in Deepfake Detection
  8. Methodology
  9. Annotating Deepfake Databases
  10. Evaluating Bias in Imbalanced Data
  11. Experimental Setup
  12. Database and Considered Attributes
  13. Deepfake Detection BackBone Models
  14. Evaluation Metrics
  15. Results
  16. ...and 10 more sections

Figures (5)

  • Figure 1: Footage of Gabon's former president Ali Bongo delivering a New Year's address, with allegations suggesting it may be a deepfake youtube.
  • Figure 2: Annotation distribution of the annotated Deepfake detection datasets - The distributions of the proposed dataset annotations are shown with the y-axis presenting percentage. For each attribute, green indicates the percentage of positive annotations, red indicates the percentage of negatively annotations, and grey represents the percentage of images that have an undefined annotation for the attribute. The distributions show that these databases are highly unbalanced concerning these attributes.
  • Figure 3: Attribute annotation correlations of the Deepfake detection databases - The 20 most positive and negative (Pearson) correlations are shown for each of the five databases. Green indicates positive correlations, while red indicates negative correlations. For working with these databases, the highly-correlating attributes should be considered to prevent misinterpretations.
  • Figure 4: Bias analysis - The relative performance $RP$ is reported with respect to the corrected relative performance $CRP$ using three Deepfake detection backbone models, EfficientNetB0 tan2019efficientnet, Xception chollet2017xception, and Capsule-Forensics-v2 nguyen1910use on four annotated databases, A-Celeb-DF, A-DFD, A-FF++, and A-DFDC. Many attributes strongly influence the detection performance.
  • Figure 5: Bias analysis on Pristine Data and Deepfake Data - The $CRP$ on the pristine data (PDRP) is reported with respect to the $CRP$ on Deepfake data (DDRP) using three Deepfake detection backbone models, EfficientNetB0 tan2019efficientnet, Xception chollet2017xception, and Capsule-Forensics-v2 nguyen1910use on our four annotated databases, A-Celeb-DF, A-DFD, A-FF++, and A-DFDC. Many of the attributes strongly influence the detection performance.