GBDF: Gender Balanced DeepFake Dataset Towards Fair DeepFake Detection

Abstract

Facial forgery by deepfakes has raised severe societal concerns. Several solutions have been proposed by the vision community to effectively combat the misinformation on the internet via automated deepfake detection systems. Recent studies have demonstrated that facial analysis-based deep learning models can discriminate based on protected attributes. For the commercial adoption and massive roll-out of the deepfake detection technology, it is vital to evaluate and understand the fairness (the absence of any prejudice or favoritism) of deepfake detectors across demographic variations such as gender and race. As the performance differential of deepfake detectors between demographic subgroups would impact millions of people of the deprived sub-group. This paper aims to evaluate the fairness of the deepfake detectors across males and females. However, existing deepfake datasets are not annotated with demographic labels to facilitate fairness analysis. To this aim, we manually annotated existing popular deepfake datasets with gender labels and evaluated the performance differential of current deepfake detectors across gender. Our analysis on the gender-labeled version of the datasets suggests (a) current deepfake datasets have skewed distribution across gender, and (b) commonly adopted deepfake detectors obtain unequal performance across gender with mostly males outperforming females. Finally, we contributed a gender-balanced and annotated deepfake dataset, GBDF, to mitigate the performance differential and to promote research and development towards fairness-aware deep fake detectors. The GBDF dataset is publicly available at: https://github.com/aakash4305/GBDF

Non-Consensual Synthetic Intimate Imagery: Prevalence, Attitudes, and Knowledge in 10 Countries

“…Like other broader AI research focusing on the issues associated with biased training data sets, deepfake detection research has identified diversity and bias problems within relevant datasets [78,84]. For example, detection techniques have been shown to be sensitive to gender, performing worse on female-based deepfakes [57] as compared to images of males. A recent systematic review of deepfake detection papers succinctly concludes that current models are sensitive to both novel or challenging conditions [74].…”

Untitled

“…Researchers in [33] manually labelled several popular datasets with gender labels. They then created a new dataset called GBDF that was gender-balanced and included gender labels.…”

“…It is important to consider newer and more innovative methods when searching for the best results, and many researchers have branched out from neural networks both to evaluate new methods entirely or to compare the results of new methods with neural network-based methods. Finally, we will survey studies that focus on the use of deepfake datasets in deepfake detection [33][34][35]. The dataset(s) used in training and testing of neural networks and other methods can have a substantial impact on results.…”