Rethinking Bias Mitigation: Fairer Architectures Make for Fairer Face Recognition

Face recognition systems are widely deployed in safety-critical applications, including law enforcement, yet they exhibit bias across a range of socio-demographic dimensions, such as gender and race. Conventional wisdom dictates that model biases arise from biased training data. As a consequence, previous works on bias mitigation largely focused on pre-processing the training data, adding penalties to prevent bias from effecting the model during training, or post-processing predictions to debias them, yet these approaches have shown limited success on hard problems such as face recognition. In our work, we discover that biases are actually inherent to neural network architectures themselves. Following this reframing, we conduct the first neural architecture search for fairness, jointly with a search for hyperparameters. Our search outputs a suite of models which Pareto-dominate all other high-performance architectures and existing bias mitigation methods in terms of accuracy and fairness, often by large margins, on the two most widely used datasets for face identification, CelebA and VGGFace2. Furthermore, these models generalize to other datasets and sensitive attributes. We release our code, models and raw data files at https://github.com/dooleys/FR-NAS.

Paper

Similar papers

Peer review

Reviewer 3GhH8/10 · confidence 4/52023-07-04

Summary

The paper aims at a fairer face recognition model. First, the authors conduct large-scale experiments to show that architectures and hyperparameters matter for fairness (Section 3). Concretely, a wide range of models in different architectures and hyperparameters are evaluated in terms of performance (metric: “Error”) and fairness (metric: Rank Disparity), showing that some models (e.g., DPN) are indeed Pareto-optimal compared to others. Motivated by this finding, unlike previous bias mitigation strategies based on a fixed neural architecture and a set of hyperparameters, the paper provides a new angle on bias mitigation by searching for fairer neural architectures and hyperparameters. The paper designs a search strategy to satisfy three desiderata: (1) both architectures and hyperparameters are optimized, (2) both accuracy and fairness are used as the objective, and (3) the searching process should be efficient. To this end, the paper uses some existing approaches, such as SMAC, Hyperband, and ParEGO. The results show that the proposed method is Pareto-optimal compared to existing methods on two face datasets. Furthermore, the experiments show that the proposed method can also generalize to other datasets and protected attributes.

Strengths

1. The proposed method is well-motivated by the experiments in Section 3 to show that architectures and hyperparameters matter for fairness. 2. The paper gives a novel angle from architectures and hyperparameters toward bias mitigation. 3. The paper provides insights into why this method works from the perspective of linear separability of protected attributes (L293). 4. The experiments are extensive. 5. In terms of the results, the proposed method is Pareto-optimal, beating existing bias mitigation methods. 6. The code is provided for better reproducibility. 7. The paper is well-written and easy to follow.

Weaknesses

## Generalization of Pareto-optimal results to different datasets I appreciate the results of cross-dataset generalization. However, Table 2 only shows the performance results. Therefore, whether or not the proposed method is still Pareto-optimal remains unknown. ### Minor Comments: The plots in Figure 2-4 are in low resolution. I suggest the authors export the plots in PDF, SVG, or EPS formats instead of image formats (e.g., jpeg). Fonts in Table 1 are in a strange aspect ratio. I suggest the authors use the “adjustbox” package to adjust the table size. L300: probes[1] -> probes [1] Appendix, L806, L808: broken \ref link to the figure. “Figure ??” -> Figure 16 Appendix, Caption of Figure 17: “(b) SMAC model second last layer (b) DPN MagFace on the second last layer” -> “(c) SMAC model second last layer (d) DPN MagFace on the second last layer.”

Questions

1. Can the authors add the fairness-performance results of cross-dataset generalization instead of only showing the performance results? 2. I appreciate the authors’ efforts in explaining why the proposed method work (L293). However, in terms of neural architecture, is there any pattern that makes some architectures more Pareto-optimal than others? This would be interesting because future works may use such a pattern to manually design a fairer architecture.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

The authors have adequately addressed the limitations (L333-348). From my perspective, the paper does not have a potential negative societal impact.

Reviewer mXr96/10 · confidence 5/52023-07-06

Summary

This paper propose a brand new framework (NAS+HPO) to mitigate biases in FR. The discussion is extensive and interesting. But experiments on authoritative face recognition dataset are required, e.g., Ms1m, Glint360 and webface260m.

Strengths

a. The presentation is easy to follow. b. The discussion is extensive and interesting. c. The paper propose a new framework (NAS+HPO jointly) to mitigate biases in FR.

Weaknesses

The reported FR performance of this method should be also verified in large scale FR datasets, since lots of methods just work in small datasets but always fail in large ones. Some authoritative face recognition datasets are suggested, e.g., Ms1m v3[1], Glint360K[2] and webface260m[3]. [1] Lightweight face recognition challenge. [2] Killing Two Birds with One Stone:Efficient and Robust Training of Face Recognition CNNs by Partial FC [3] WebFace260M: A Benchmark Unveiling the Power of Million-Scale Deep Face Recognition

Questions

null

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

3 good

Contribution

4 excellent

Limitations

null

Reviewer Nw7v8/10 · confidence 4/52023-07-06

Summary

The paper focuses on Bias Mitigation for face identification, i.e., ensuring that face identification works “well” for different identities: gender, race, etc. Unlike prior work, which focuses on model backbone agnostic methods to mitigate bias, this work explores the relevance of the inductive bias encoded in different Deep Learning model's backbones to the bias issue. In other words, are there model backbones better at learning robust features and ensuring good performance on samples from different identities? The authors uncover model configurations that significantly improve performance over standard model backbones through an extensive empirical analysis based on Neural Architecture Search on two face identification datasets. More surprisingly, the authors prove that these backbones' performance is better or more competitive than when paired with standard bias mitigation methods. Moreover, the authors confirm the generalizability of the configurations uncovered using CelebA and VGGFace2 by testing them on other datasets, further confirming their competitive performance. Finally, the authors analyze the new model configurations and ensure that the features learned by the models are less likely to be discriminative between the biased groups confirming that they learn more diverse and robust features.

Strengths

1- The work is well motivated; prior work needs to include an analysis of model architecture relevance to the bias issue. 2-The results are interesting and relevant to fairness/bias community practitioners. 3-The advantages of newly discovered model configurations are explored by a well-designed empirical analysis that confirms the configurations' more robust learned features.

Weaknesses

1- Some choices in the experimental design would benefit from further motivation. For example: Why was the “multi-fidelity Bayesian optimization method SMAC3” chosen in particular? Are there other methods that could also work? 2- SMAC_301 was the architecture that works well across datasets. I understand that 301 denotes the operations that constitute the novel architecture. However, the authors do not discuss the details of these operations or why they think they are meaningful choices compared to other choices ruled out by the NPS. Some discussion here would be helpful. Why do the authors think this configuration is better able to learn non-linearly separable bias features? 3- In Section 3, why are the models trained on a gender-balanced subset of the dataset? Wouldn’t one want to train on a gender-imbalanced split to see which architectures are less likely to be influenced by the imbalance? Is this the same in Section 4? Please clarify. 4- In the analysis in Section 4.2, were the hyper-parameters (learning rate, optimizer) of the timm models tuned too? Or were they the ones used by the original papers? If it is the latter, then it is an unfair comparison to the SMAC models since those hyper-parameters were tuned as outlined in Section 4.1. 5- From reading the paper in detail, I understand now that one of the motivations of Section 3 analysis was to limit the number of architectures explored in Section 4 (only DPN was considered since it achieved Pareto optimal performance on both datasets). However, I did not get that from the first read, so further clarification in the text about that would be helpful. 6- In Section 3, the hyper-parameters of the different models were not optimized. It is likely very computationally expensive. Nevertheless, it does undermine some elements of the analysis, so I would make that clear as limitations.

Questions

I am overall positive about this work. However, I need further clarifications as outlined in the Weakness sections, particularly questions: (2,3,4). I am happy to increase my score upon adequate further clarification.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

I commend the authors for explicitly discussing their work's technical limitations and that while it improves the notion of technical fairness, the advancement could still be harmful in downstream applications.

Reviewer Kze16/10 · confidence 4/52023-07-07

Summary

This paper presents a new perspective on bias mitigation in machine learning models, challenging the conventional belief that one should first find the highest-performing model and then apply a bias mitigation strategy. The authors propose that finding a fairer architecture offers significant gains compared to conventional bias mitigation strategies. To test this hypothesis, they conduct the first neural architecture search for fairness and a search for hyperparameters in face recognition.

Strengths

This paper proposes a new way to mitigate biases in face recognition systems from the perspective of fairer model architectures. This paper conducts the first large-scale analysis of the impact of architectures and hyperparameters on bias in face recognition, demonstrating that the implicit convention of choosing the highest-accuracy architectures is a sub-optimal strategy for fairness. This paper may be the first to apply existing tools from NAS and HPO to design a fair face recognition model automatically.

Weaknesses

(1)According to the definition of rank difference, the smaller the model's error, the fairer the model. Therefore, the difference in model parameters will lead to a difference in model performance. There is a lack of consistent constraints on the magnitude of the model parameters when searching for a fairer structure in this paper. (2)Why the results in Table 2 are far worse than those reported in other papers.

Questions

Please refer to the WeakNess

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Yes

Reviewer jfj27/10 · confidence 4/52023-07-26

Summary

The authors offer a fresh view on mitigating fairness bias in ML by leveraging neural architecture (NAS) search and hyperparameter optimization (HPO). The authors demonstrate their idea on the exemplary problem of face identification, where fairness biases have tangible consequences on society. They utilize a wide range of model architectures, and define a NAS+HPO search strategy where multi-objective optimization helps balance the tradeoff between accuracy and fairness (as quantified e.g. by the rank disparity metric) .

Strengths

+ The authors formally describe a new paradigm compared with classical bias mitigation techniques that have traditionally focused on postprocessing/rectifying ML predictions, preprocessing/balancing the datasets, or extending the loss. + The solution is systematically devised and the experiments are straightforward to follow. + The experimental results are insightful and helpful in practice. I find it inspiring that the NAS models generalize to new protected attributes in new datasets. + The authors provide their source code besides a variety of analysis scenarios available to run via notebooks.

Weaknesses

- There was no discussion on the impact of pretraining. With the availability of a large number of foundational models, it would be very relevant to shed light into which ones generalize better and why. - The theoretical analysis is a. bit lacking. I was expecting more explanation of why the NAS models outperform other bias mitigation strategies and better generalize to other sensitive attributes. Is it because you are forcing the model to work harder and to avoid misusing these sensitive attributes as shortcuts when making predictions? (this would explain the reduced linear separability of protected attributes). - The visualization of the results could be more insightful. For example a confusion matrix / similarity matrices might be useful to conduct error analysis and shed light into the improvements facilitated by the NAS paradigm. [Such analysis](https://arxiv.org/abs/2007.06068) has revealed many characteristics of VGGFace, e.g., oftentimes, gender misclassification is due to labeling issue (e.g. the image scrapped is not of the actor, but of their opposite-gender spouse). Minor and language issues: - The figures could be better annotated (e.g. to explain that the two red dots in Figure 2 are two variants of DPN) - when comparing to other bias mitigation techniques => compared? - at the most extreme low errors => unclear - oepration - to supports - are Pareto-optimal the top performing … => are the Pareto-optimal top-performing … - recognititon

Questions

Have you considered other alternatives to SMAC3 or ParEGO? How generalizable are the insights in section 4.2. to possible alternatives?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

4 excellent

Limitations

Sufficiently discussed

Reviewer 3GhH2023-08-13

I have read the authors' responses and reviewers' comments. The response addresses my concern. I raise my rating to "Strong Accept." I encourage the authors to add the response to the final version.

Reviewer Nw7v2023-08-13

The response addresses my concerns. I encourage the authors to revise the manuscript and include the updates. I revised my score accordingly.

Reviewer Kze12023-08-14

Response

I have read the authors' responses and reviewers' comments. The response addresses my concern. I keep my rating.

Reviewer jfj22023-08-14

I would like to thank the authors for their detailed response and for conducting the additional experiments. This confirmed my assessment about the utility and the solid results of the presented method.

Program Chairsdecision2023-09-21

Decision

Accept (oral)

© 2026 NYSGPT2525 LLC