Deep learning models in computer vision, such as Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), have been found to exhibit significant biases related to factors such as gender and ethnicity. These biases often originate from inherent imbalances in training data predominantly sourced from the internet. In this study, we aim to address gender bias in computer vision models by curating a specialized dataset that highlights gender-related disparities. Additionally, we measure dataset diversity across six datasets (FFHQ, WIKI, IMDB, LFW, UTK Faces, diverse dataset), five professions (CEO, engineer, nurse, politician, and teacher) and diferent query retrieval tasks using the Image Similarity Score (ISS). To reduce learned gender biases and increase data diversity, we propose adversarial data augmentation techniques that specifically target facial regions within images. These techniques, named Partial Mix (PM), that partially mixes two gendered faces in a squared pattern, and Noise Addition (NA), that adds noise to the facial region, are designed to mitigate bias. Our experimental results demonstrate increased data diversity across the six datasets and professions, along with reduction in gender bias for CNN-based models. However, these adversarial techniques were less efective in reducing bias for Vision Transformers. This discrepancy highlights the unique challenges for bias mitigation posed by ViTs. Consistent with prior research, our findings indicate that ViTs learn from a broader set of visual cues compared to CNNs. This increased sensitivity makes ViTs more prone to amplifying biases, emphasizing the need for tailored bias mitigation strategies when deploying these models in real-world applications.
Social biases related to ethnicity [
Numerous strategies have been proposed to mitigate bias in computer vision models. These include
the expansion of dataset diversity, as outlined in Kärkkäinen et al.’s work [
The remainder of this paper is organized as follows: Section 2 briefly reviews existing related work, Section 3 explains the proposed methodology, Section 4 discusses the experimental setup, the insight and findings and finally Section 5 presents the conclusions.
(a) without Data augmentation
(b) with Data augmentation(Ours) (c) without Data augmentation (d) with Data augmentation(Ours)
The issue of gender bias in computer vision models has received significant attention within the research community, with a multitude of proposed techniques for mitigating this bias. These approaches encompass various strategies, including the manipulation of learned representations [11], adjustments to the training dataset [12], and the application of adversarial methods [13]. It is important to note that a majority of these debiasing techniques have been tailored for Convolutional Neural Networks (CNNs). However, as the landscape of computer vision continues to evolve, Vision Transformers (ViTs) have gained prominence, often surpassing CNNs in numerous tasks, such as image classification [ 14, 15]. Mandal et al. observed that ViTs tend to exacerbate social biases to a greater extent when compared to CNNs [16].
Data augmentation aims to increase data diversity so that deep learning models can be trained to improve the generalization ability [17, 18, 19]. At present only limited data augmentation work has focused on debiasing. Zhang et al. [7] explored machine learning fairness in image classification, addressing bias from imbalanced data and harnessing adversarial examples as data augmentation for data distribution balance. Li et al. [9], aim to improve cross-bias generalization using data augmentation. They introduce “safety” and “unbiasedness” constraints to address the influence of biased cues in training data without manual intervention. Smith et al. [8], tackles gender and age classification biases by leveraging data augmentation techniques. The authors introduce an innovative approach that optimizes data augmentation settings through an evolutionary process, efectively reducing bias and improving model generalization. Though these above research works explore and mitigate gender bias using diferent data augmentation techniques, in our work we introduce two novel adversarial data augmentation techniques to address gender bias. The efect of data augmentation on gender debiasing is illustrated in Figure 1
In this section, we introduce an alternative methodology. Initially, we employ facial recognition on the input image using the well-established and highly eficient face recognition algorithm, Single-shot Detection (SSD) [20]. To perform this task, we utilize a pre-trained model 1 and detected faces using OpenCV 2. Once the facial region has been successfully detected within the original image, , we proceed to apply the newly proposed data augmentation techniques as follows: 1. Partial Mixing (PM) : In this approach, the facial regions and of male and female, respectively are taken. Each is divided into four equal parts, and a random selection of squares from both facial regions is mixed. A mask, , is partitioned into four segments, each filled with either 0’s or 1’s to respectively include or exclude those squares. Subsequently, an element-wise multiplication is conducted between the mask, , and the male facial region, , and 1 − and female facial region, , then both are added, resulting in the generation of the augmented image, ˜, as illustrated in Equation 1. Finally, the augmented facial region ˜ is reinserted into the original image. The overall process is depicted in Figure 2.
x˜a = ⊙ xm + (1 − ) ⊙ xf 2. Noise addition (NA): In this strategy, we incorporate uniformly distributed noise, generated within the range of 0 to 1, as expressed in Equation 2. This randomly generated noise, denoted as , is then added to the facial region, or . Consequently, an augmented facial region, ˜, is produced, as outlined in Equation 3.
= (0, 1) ˜ = + (1) (2) (3) Then ˜ is placed back to its position in the original images. The overall process is shown in Figure 3. 1https://raw.githubusercontent.com/opencv/opencv_3rdparty/dnn_samples_face_detector_20180205_fp16/res10_300x300_ssd_iter_140000_fp16.c 2https://docs.opencv.org/3.4/d6/d0f/group__dnn.html
Image Similarity Score (ISS) is used to measure the diversity of a dataset. There are two variants of ISS – (i) ISS, which measures diversity in the dataset, (ii) ISS, which measures diversity across the datasets – both introduced by Mandel [4] and both with a range of 0 to 2. To measure ISS, we used six diverse datasets (FFHQ [21] , WIKI [22] , IMDB [22], LFW [23], UTK [24] and Diverse Dataset [4] ), ifve professions and a query retrieval task dataset, as described in Mandel [ 4]. The Image Similarity Score (ISS) measures how similar two images are based on features extracted by a pre-trained Convolutional Neural Network (CNN). In this study, we use VGG16, a 16-layer deep CNN trained on the ImageNet dataset. The feature extraction layers of VGG16 were employed to capture features from the images. To reduce the dimensionality of these extracted features, we applied Principal Component Analysis (PCA). For two images, 1 and 2, with corresponding feature vectors 1 and 2, the similarity between the images is calculated as shown in Equation 4 [4] and also algorithm for and are proposed by Mandal et al. [4]. A higher Image Similarity Score indicates greater dissimilarity between images. sim (1, 2) = 1 −
We curated a visual dataset with ten classes: CEO, Engineer, Baseball, Rugby, Snowboarding, Nurse,
School Teacher, Hairdryer, Shopping, and Dollhouse. The first five categories are generally (in a social or
stereotypical sense) male-dominated and the last five female are female-dominated [
The results in Table 1 demonstrate the efectiveness of both of our methods, “with PM” and “with NA”, in improving the Intra-dataset Image Similarity Score (ISS) across multiple datasets. Our PM approach consistently outperforms the baseline for all datasets, achieving the highest ISS values representing greater diversity in the results. Specifically, the PM method achieves notable improvements on the FFHQ, Diverse Dataset, WIKI, IMDB, LFW, and UTK datasets, with the highest ISS observed for the IMDB dataset at 1.21. In contrast, the NA approach, while still improving upon the baseline, yields slightly lower scores compared to the PM method but consistently surpasses the baseline. This demonstrates that both methods contribute to improved dataset diversity, with the PM approach being more efective overall.
the baseline (0.8990). A similar trend is observed for the Engineer query, where the NA method outperforms PM in most regions, especially for Russian-East Europe and Spanish-Latin America, where NA achieves ISS values of 1.0076 and 0.9981, respectively. For the Nurse query, the NA method again consistently outperforms both the baseline and PM, with a remarkable improvement in the Swahili-Sub Saharan Africa region, where the ISS increases from 0.9585 (baseline) to 0.9955. The Politician query also shows substantial gains with both approaches, particularly in Russian-East Europe, where the NA method reaches an ISS of 0.9982, an increase over the baseline of 0.9383. Finally, for the School Teacher query, both methods show increased performance in nearly all regions, with the NA method showing slightly higher ISS scores, particularly in Arabic-West Asia & North Africa (1.0155) and Spanish-Latin America (0.9982).
As presented in Table 3, both approaches, PM and NA, consistently outperform the baseline in both ISS and ISS evaluations. For the CEO query, the cross-dataset score (ISS) for the NA method is slightly higher (0.9960) compared to PM (0.9956), showing a marginal improvement over the baseline. A similar pattern is observed for the Engineer and Politician queries, where the NA method again shows higher cross-dataset performance. For Nurse and School Teacher, the PM method performs slightly better in ISS, but the NA method maintains higher cross-dataset scores. This is particularly evident for the School Teacher query, where the NA method scores 1.001 in ISS and 0.9931 in ISS, outperforming both the baseline and PM.
When averaged across all queries, the PM method achieves the highest mean ISS score (0.9958), while the NA method follows closely with 0.9944, both outperforming the baseline (0.9803). Similarly, for ISS, PM leads with 0.9966, followed closely by NA (0.9957), both again surpassing the baseline value of 0.9885. These results emphasize the efectiveness of our methods, particularly the PM approach, in increasing dataset diversity both within and across datasets.
As shown in Table 4, CNN models demonstrated improvements in accuracy with the application of the Uniform Noise Blur technique, and in two cases (Inception V3 and Xception) with Partial Mix, though none surpassed the performance of manually debiased data. Inception V3 and Xception showed the most consistent gains, while VGG16 saw only minor improvement. In contrast, Vision Transformers (ViTs) showed no improvements with either augmentation method, indicating that these techniques were inefective in reducing bias for ViTs. This highlights the need for more tailored approaches for bias mitigation in ViTs, which likely depend on cues beyond facial features.
Our study demonstrates that adversarial data augmentation techniques, Partial Mix (PM) and Noise Addition (NA), significantly enhance dataset diversity and reduce gender bias, particularly in CNN models. This is reflected in improvements across both Intra-dataset and Cross-dataset Image Similarity Scores (ISS), indicating a more visually diverse and balanced representation of gender-related features. CNN models trained with these augmented datasets show a noticeable reduction in bias, supporting the efectiveness of targeted facial-region augmentation. However, Vision Transformers (ViTs) do not exhibit the same reduction in gender bias. Despite the use of the same augmentation techniques, ViTs continue to amplify biases, likely due to their ability to learn from broader visual cues, such as clothing and objects, beyond just facial features. This heightened sensitivity makes them more resistant to bias mitigation through facial-focused augmentations, leading to less improvement in diversity and fairness compared to CNNs. In summary, while adversarial data augmentation enhances diversity and mitigates bias in CNNs, it is less efective for ViTs, which require more comprehensive strategies that account for the broader context in which biases are learned. Future work should focus on developing bias mitigation methods that target a wider range of visual signals, particularly for ViTs, to ensure equitable and fair representation in computer vision models.
This research was supported by Science Foundation Ireland under grant numbers 18/CRT/6223 (SFI Centre for Research Training in Artificial intelligence), SFI/12/RC/2289/ _2 (Insight SFI Research Centre for Data Analytics), 13/RC/2094/ _2 (Lero SFI Centre for Software) at Dublin City University and 13/RC/2106/ _2 (ADAPT SFI Research Centre for AI-Driven Digital Content Technology). For the purpose of Open Access, the author has applied a CC BY public copyright licence to any Author Accepted Manuscript version arising from this submission. [4] A. Mandal, S. Leavy, S. Little, Dataset diversity: Measuring and mitigating geographical bias in image search and retrieval, in: Proceedings Of The 1st International Workshop On Trustworthy AI For Multimedia Computing, 2021, pp. 19–25. [5] A. Mandal, S. Little, S. Leavy, Gender bias in multimodal models: A transnational feminist approach considering geographical region and culture, ArXiv Preprint ArXiv:2309.04997 (2023). [6] A. Mandal, S. Leavy, S. Little, Measuring bias in multimodal models: Multimodal composite association score, in: International Workshop On Algorithmic Bias In Search And Recommendation, 2023, pp. 17–30. [7] Y. Zhang, J. Sang, Towards accuracy-fairness paradox: Adversarial example-based data augmentation for visual debiasing, in: Proceedings Of The 28th ACM International Conference On Multimedia, 2020, pp. 4346–4354. [8] P. Smith, K. Ricanek, Mitigating algorithmic bias: Evolving an augmentation policy that is nonbiasing, in: Proceedings Of The IEEE/CVF Winter Conference On Applications Of Computer Vision Workshops, 2020, pp. 90–97. [9] L. Li, F. Tang, J. Cao, X. Li, D. Wang, Bias oriented unbiased data augmentation for cross-bias representation learning, Multimedia Systems 29 (2023) 725–738. [10] F. Chollet, Xception: Deep learning with depthwise separable convolutions, in: Proceedings Of
The IEEE Conference On Computer Vision And Pattern Recognition, 2017, pp. 1251–1258. [11] Z. Wang, K. Qinami, I. Karakozis, K. Genova, P. Nair, K. Hata, O. Russakovsky, Towards fairness in visual recognition: Efective strategies for bias mitigation, in: Proceedings Of The IEEE/CVF Conference On Computer Vision And Pattern Recognition, 2020, pp. 8919–8928. [12] C. Schumann, S. Ricco, U. Prabhu, V. Ferrari, C. Pantofaru, A step toward more inclusive people annotations for fairness, in: Proceedings Of The 2021 AAAI/ACM Conference On AI, Ethics, And Society, 2021, pp. 916–925. [13] T. Wang, J. Zhao, M. Yatskar, K.-W. Chang, V. Ordonez, Balanced datasets are not enough: Estimating and mitigating gender bias in deep image representations, in: Proceedings Of The IEEE/CVF International Conference On Computer Vision, 2019, pp. 5310–5319. [14] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, Others, An image is worth 16x16 words: Transformers for image recognition at scale, ArXiv Preprint ArXiv:2010.11929 (2020). [15] S. Khan, M. Naseer, M. Hayat, S. Zamir, F. Khan, M. Shah, Transformers in vision: A survey, ACM
Computing Surveys (CSUR) 54 (2022) 1–41. [16] A. Mandal, S. Leavy, S. Little, Biased attention: Do vision transformers amplify gender bias more than convolutional neural networks?, ArXiv Preprint ArXiv:2309.08760 (2023). [17] C. Shorten, T. M. Khoshgoftaar, A survey on image data augmentation for deep learning, Journal of big data 6 (2019) 1–48. [18] T. Kumar, M. Turab, K. Raj, A. Mileo, R. Brennan, M. Bendechache, Advanced data augmentation approaches: A comprehensive survey and future directions, ArXiv Preprint ArXiv:2301.02830 (2023). [19] T. Kumar, A. Mileo, R. Brennan, M. Bendechache, Rsmda: Random slices mixing data augmentation,
Applied Sciences 13 (2023) 1711. [20] C.-Y. Fu, W. Liu, A. Ranga, A. Tyagi, A. Berg, Dssd: Deconvolutional single shot detector, ArXiv
Preprint ArXiv:1701.06659 (2017). [21] T. Karras, S. Laine, T. Aila, Nvlabs/fhq-dataset, 2019. URL: https://github.com/NVlabs/fhq-dataset. [22] R. Rothe, R. Timofte, L. Van Gool, Dex: Deep expectation of apparent age from a single image, in: Proceedings of the IEEE International Conference on Computer Vision Workshops, 2015, pp. 10–15. [23] E. Learned-Miller, G. Huang, A. RoyChowdhury, H. Li, G. Hua, Labeled faces in the wild: A survey, in: Advances in Face Detection and Facial Image Analysis, 2016, pp. 189–248. [24] Z. Zhang, Y. Song, H. Qi, Age progression/regression by conditional adversarial autoencoder, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017.