pith. sign in

arxiv: 2605.19549 · v1 · pith:6BAXR6EFnew · submitted 2026-05-19 · 💻 cs.SE · cs.LG

Provable Fairness Repair for Deep Neural Networks

Jianan Ma , Jingyi Wang , Qi Xuan , Zhen Wang This is my paper

Pith reviewed 2026-05-20 04:39 UTC · model grok-4.3

classification 💻 cs.SE cs.LG
keywords fairness repairdeep neural networksprovable guaranteesinterval bound propagationMILPmodel repairindividual discriminationethical AI
0
0 comments X

The pith

ProF repairs deep neural networks for fairness with provable guarantees by using interval bound propagation to guide a MILP-based adjustment that ensures consistent outputs over neighborhoods of biased inputs.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces ProF, a framework for repairing deep neural networks to remove individual discrimination while offering mathematical guarantees that the fix works across similar inputs. It works by applying interval bound propagation to calculate safe ranges for the network's behavior in a region around any example that shows bias. These ranges then shape a set of constraints that get turned into a mixed-integer linear program, which solvers can handle to find the necessary model changes. A sympathetic reader would care because this shifts fairness fixes from trial-and-error on data to methods that prove the bias is gone for whole groups of related cases, including ones not seen in training. The evaluations show the repaired models generalize fairness to most of the dataset and input space.

Core claim

The central claim is that integrating fairness constraints with model modifications into a unified constraint-solving formulation, transformed into a Mixed-Integer Linear Programming problem, allows the solution to induce a repaired model with guaranteed fairness over the whole set S(x) around biased samples, where the guarantees come from the sound capture of outputs via interval bound propagation.

What carries the argument

The central mechanism is the use of interval bound propagation to derive bounds on model outputs over the set S(x) of inputs around a biased sample, which then guide the formulation and solution of a Mixed-Integer Linear Programming problem to repair the model for consistent outputs on that set.

If this is right

  • The repaired model produces consistent outputs for all inputs in S(x) for each biased sample processed.
  • Provable fairness repair generalizes to up to 95.93 percent on full datasets and 93.16 percent on the entire input space.
  • The framework supports multiple sensitive attributes and various practical fairness definitions with the same guarantees.
  • Around 90 percent fairness improvement is achieved alongside the provable aspects.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If interval bound propagation can be applied to other properties, similar repair techniques might address robustness or other ethical concerns in neural networks.
  • Scaling the MILP solver or using approximate methods could make this repair approach feasible for larger models in practice.
  • The reliance on neighborhoods suggests potential connections to adversarial robustness techniques that also use bound propagation.

Load-bearing premise

Interval bound propagation must soundly and tightly capture the model outputs over the entire set S(x) around each biased sample so that the derived bounds correctly guide the repair to enforce fairness everywhere in that set.

What would settle it

Finding a biased sample x where after applying the MILP solution the repaired model still gives different outputs for two inputs in S(x), or observing generalization rates much lower than 95 percent on the test data.

Figures

Figures reproduced from arXiv: 2605.19549 by Jianan Ma, Jingyi Wang, Qi Xuan, Zhen Wang.

Figure 1
Figure 1. Figure 1: Interval bound propagation on an example NN. All [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overview of PROF framework. For notational simplicity, we use f1:L to denote fL−1 ◦ · · · ◦ f1 hereafter. on given IDI pairs followed by parameter modification, or (2) retraining the models to minimize the distances between features extracted from the IDI pairs. In other words, they aim to reduce the distances between specific feature pairs (discrete point pairs in the purple region in [PITH_FULL_IMAG… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of repair with concrete bounds and sym [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Loss dynamics during the bounds tightening in Step 1 [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison between the naive repair method and [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Full results of loss dynamics during the progressive bounds tightening process in Step 1 of P [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
read the original abstract

Deep neural networks (DNNs) are suffering from ethical issues such as individual discrimination. In response, extensive NN repair techniques have been developed to adjust models and mitigate such undesired behaviors. However, existing fairness repair methods are typically data-centric, which often lack provable guarantees and generalization to unseen samples. To overcome these limitations, we propose ProF, a novel fairness repair framework with provable guarantees. The key intuition of ProF is to leverage interval bound propagation (a widely used NN verification technique) to soundly capture model outputs over the whole set $S(\mathbf{x})$ around a biased sample $\mathbf{x}$. The derived bounds are utilized to guide fairness repair which encourages the model to produce consistent outputs on $S(\mathbf{x})$. Specifically, we integrate fairness constraints and model modifications into a unified constraint-solving formulation, which can be transformed to a Mixed-Integer Linear Programming (MILP) problem solvable by off-the-shelf solvers. The solution to the MILP problem effectively induces a repaired model with guaranteed fairness over the whole set $S(\mathbf{x})$. We evaluate ProF on four widely used benchmark datasets and demonstrate that it achieves provable fairness repair, with generalization of up to 95.93\% on full datasets and 93.16\% on the entire input space. Notably, ProF can be easily configured to support multiple sensitive attributes and more practical fairness definitions, while providing provable repair guarantees and delivering around 90\% fairness improvement. Our code is available at https://github.com/nninjn/ProF.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes ProF, a framework for provable fairness repair of DNNs. It uses interval bound propagation (IBP) to compute sound output bounds over a neighborhood set S(x) around a biased sample x, then encodes fairness constraints together with model modifications into a single constraint problem that is cast as a MILP. Solving the MILP is claimed to produce a repaired network whose outputs are guaranteed consistent (hence fair) on every point in S(x). Experiments on four benchmarks report up to 95.93% generalization on full datasets, 93.16% on the entire input space, and roughly 90% fairness improvement, with support for multiple sensitive attributes.

Significance. A sound realization of the approach would constitute a meaningful advance: it replaces purely data-driven repair with a verification-guided, constraint-based method that supplies formal guarantees over continuous neighborhoods rather than finite samples. The public code release and the ability to handle multiple fairness definitions are additional strengths that aid reproducibility and practical adoption.

major comments (2)
  1. [§3] §3 (MILP formulation of IBP with variable weights): the encoding must handle products of the form W · [L,U] where both W and the interval bounds are decision variables. The manuscript should state explicitly whether the formulation uses an exact linearization, McCormick envelopes, or a restriction (e.g., bias-only updates). If a relaxation is employed, a proof is required that any feasible MILP solution still implies the repaired network satisfies the fairness property on the whole set S(x); otherwise the central guarantee does not follow.
  2. [§4.2] §4.2 (generalization claims): the reported 93.16% coverage of the entire input space is derived from neighborhoods around a finite set of biased samples. The paper must clarify how this percentage is computed and whether it constitutes a formal guarantee over the full domain or only an empirical estimate on held-out data.
minor comments (2)
  1. [§2] Notation for the neighborhood radius and the set S(x) should be introduced once in §2 and used consistently thereafter.
  2. [Table 2] Table 2: the column headers for the different fairness metrics are not fully aligned with the definitions given in §2.3; add a footnote or reference.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and the recommendation for major revision. We address each major comment point by point below, indicating the revisions we plan to incorporate.

read point-by-point responses

  1. Referee: [§3] §3 (MILP formulation of IBP with variable weights): the encoding must handle products of the form W · [L,U] where both W and the interval bounds are decision variables. The manuscript should state explicitly whether the formulation uses an exact linearization, McCormick envelopes, or a restriction (e.g., bias-only updates). If a relaxation is employed, a proof is required that any feasible MILP solution still implies the repaired network satisfies the fairness property on the whole set S(x); otherwise the central guarantee does not follow.

    Authors: We thank the referee for highlighting this important detail in the MILP encoding. In the current formulation, we restrict modifications to the bias terms while keeping all weight matrices fixed as constants. This restriction permits an exact linearization of the IBP bounds without requiring McCormick envelopes or other relaxations. We will revise §3 to state this restriction explicitly, include the full linearization equations, and add a short argument confirming that any feasible MILP solution preserves the soundness of the IBP bounds and therefore the fairness guarantee over every point in S(x). revision: yes

  2. Referee: [§4.2] §4.2 (generalization claims): the reported 93.16% coverage of the entire input space is derived from neighborhoods around a finite set of biased samples. The paper must clarify how this percentage is computed and whether it constitutes a formal guarantee over the full domain or only an empirical estimate on held-out data.

    Authors: We agree that the 93.16% figure requires clarification. This percentage is obtained by discretizing the input domain and measuring the fraction of sampled points that lie inside the union of the neighborhoods S(x) around the selected biased samples; it is therefore an empirical coverage estimate on the benchmark domains rather than a formal guarantee over the entire continuous input space. We will revise §4.2 to describe the discretization and sampling procedure in detail and to qualify the reported figure as an empirical estimate. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the derivation chain

full rationale

The paper derives its central claim by directly encoding IBP bounds, fairness constraints, and model modification variables into a MILP formulation whose solution is asserted to induce a repaired network satisfying the encoded properties on S(x). This is a constraint-solving construction rather than a fitted parameter renamed as a prediction, a self-definitional loop, or a load-bearing self-citation chain. The soundness is claimed to rest on the external properties of IBP and MILP solvers, not on quantities defined inside the present work or its authors' prior results. No equations or steps in the abstract or description reduce the guarantee to an input by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework depends on the soundness of interval bound propagation for neural networks and the existence of a feasible MILP solution that preserves model utility while enforcing fairness; no new entities are postulated.

free parameters (1)
  • neighborhood radius for S(x)
    Defines the size of the input set around each biased sample; its value is chosen to balance coverage and computational cost.
axioms (1)
  • domain assumption Interval bound propagation provides sound over-approximations of neural network output ranges over compact input sets.
    Invoked to derive bounds that guide the fairness constraints in the MILP formulation.

pith-pipeline@v0.9.0 · 5807 in / 1260 out tokens · 40732 ms · 2026-05-20T04:39:17.839403+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

62 extracted references · 62 canonical work pages · 1 internal anchor

  1. [1]

    Application of deep learning for retinal image analysis: A review,

    M. Badar, M. Haris, and A. Fatima, “Application of deep learning for retinal image analysis: A review,”Computer Science Review, vol. 35, p. 100203, 2020

    work page 2020

  2. [2]

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,”arXiv preprint arXiv:1810.04805, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  3. [3]

    Deep learning-based autonomous driving systems: A survey of attacks and defenses,

    Y . Deng, T. Zhang, G. Lou, X. Zheng, J. Jin, and Q.-L. Han, “Deep learning-based autonomous driving systems: A survey of attacks and defenses,”IEEE Transactions on Industrial Informatics, vol. 17, no. 12, pp. 7897–7912, 2021

    work page 2021

  4. [4]

    Evaluating the predictive validity of the compas risk and needs assessment system,

    T. Brennan, W. Dieterich, and B. Ehret, “Evaluating the predictive validity of the compas risk and needs assessment system,”Criminal Justice and behavior, vol. 36, no. 1, pp. 21–40, 2009

    work page 2009

  5. [5]

    Empirical observation of negligible fairness–accuracy trade-offs in machine learning for public policy,

    K. T. Rodolfa, H. Lamba, and R. Ghani, “Empirical observation of negligible fairness–accuracy trade-offs in machine learning for public policy,”Nature Machine Intelligence, vol. 3, no. 10, pp. 896–904, 2021

    work page 2021

  6. [6]

    Consumer credit-risk mod- els via machine-learning algorithms,

    A. E. Khandani, A. J. Kim, and A. W. Lo, “Consumer credit-risk mod- els via machine-learning algorithms,”Journal of Banking & Finance, vol. 34, no. 11, pp. 2767–2787, 2010

    work page 2010

  7. [7]

    Black box fairness testing of machine learning models,

    A. Aggarwal, P. Lohia, S. Nagar, K. Dey, and D. Saha, “Black box fairness testing of machine learning models,” inProceedings of the 2019 27th ACM joint meeting on european software engineering conference and symposium on the foundations of software engineering, 2019, pp. 625–635

    work page 2019

  8. [8]

    White-box fairness testing through adversarial sampling,

    P. Zhang, J. Wang, J. Sun, G. Dong, X. Wang, X. Wang, J. S. Dong, and T. Dai, “White-box fairness testing through adversarial sampling,” inProceedings of the ACM/IEEE 42nd international conference on software engineering, 2020, pp. 949–960

    work page 2020

  9. [9]

    Dissecting global search: A simple yet effective method to boost individual discrimination testing and repair,

    L. Quan, T. Li, X. Xie, Z. Chen, S. Chen, L. Jiang, and X. Li, “Dissecting global search: A simple yet effective method to boost individual discrimination testing and repair,” in2025 IEEE/ACM 47th International Conference on Software Engineering (ICSE). IEEE Computer Society, 2025, pp. 771–771

    work page 2025

  10. [10]

    Maft: Efficient model-agnostic fairness testing for deep neural networks via zero-order gradient search,

    Z. Wang, M. Zhang, J. Yang, B. Shao, and M. Zhang, “Maft: Efficient model-agnostic fairness testing for deep neural networks via zero-order gradient search,” inProceedings of the IEEE/ACM 46th International Conference on Software Engineering, 2024, pp. 1–12

    work page 2024

  11. [11]

    Efficient white-box fairness testing through gradient search,

    L. Zhang, Y . Zhang, and M. Zhang, “Efficient white-box fairness testing through gradient search,” inProceedings of the 30th ACM SIGSOFT International Symposium on Software Testing and Analysis, 2021, pp. 103–114

    work page 2021

  12. [12]

    Neuronfair: Interpretable white-box fairness testing through biased neuron identification,

    H. Zheng, Z. Chen, T. Du, X. Zhang, Y . Cheng, S. Ji, J. Wang, Y . Yu, and J. Chen, “Neuronfair: Interpretable white-box fairness testing through biased neuron identification,” inProceedings of the 44th International Conference on Software Engineering, 2022, pp. 1519–1531

    work page 2022

  13. [13]

    Fairify: Fairness verification of neural net- works,

    S. Biswas and H. Rajan, “Fairify: Fairness verification of neural net- works,” in2023 IEEE/ACM 45th International Conference on Software Engineering (ICSE). IEEE, 2023, pp. 1546–1558

    work page 2023

  14. [14]

    Fairquant: Certifying and quan- tifying fairness of deep neural networks,

    B. H. Kim, J. Wang, and C. Wang, “Fairquant: Certifying and quan- tifying fairness of deep neural networks,” in2025 IEEE/ACM 47th International Conference on Software Engineering (ICSE). IEEE Computer Society, 2024, pp. 191–203

    work page 2024

  15. [15]

    Probabilistic verification of neural networks against group fairness,

    B. Sun, J. Sun, T. Dai, and L. Zhang, “Probabilistic verification of neural networks against group fairness,” inFormal methods: 24th international symposium, FM 2021, virtual event, November 20–26, 2021, proceedings

    work page 2021

  16. [16]

    Springer, 2021, pp. 83–102

    work page 2021

  17. [17]

    Causality-based neural network repair,

    B. Sun, J. Sun, L. H. Pham, and J. Shi, “Causality-based neural network repair,” inProceedings of the 44th International Conference on Software Engineering, 2022, pp. 338–349

    work page 2022

  18. [18]

    Runner: Responsible unfair neuron repair for enhancing deep neural network fairness,

    T. Li, Y . Cao, J. Zhang, S. Zhao, Y . Huang, A. Liu, Q. Guo, and Y . Liu, “Runner: Responsible unfair neuron repair for enhancing deep neural network fairness,” inProceedings of the 46th IEEE/ACM International Conference on Software Engineering, 2024, pp. 1–13

    work page 2024

  19. [19]

    Isolation-based debugging for neural networks,

    J. Chen, J. Wang, Y . Sun, P. Cheng, and J. Chen, “Isolation-based debugging for neural networks,” inProceedings of the 33rd ACM SIGSOFT International Symposium on Software Testing and Analysis, 2024, pp. 338–349

    work page 2024

  20. [20]

    Neufair: Neural network fairness repair with dropout,

    V . A. Dasu, A. Kumar, S. Tizpaz-Niari, and G. Tan, “Neufair: Neural network fairness repair with dropout,” inProceedings of the 33rd ACM SIGSOFT International Symposium on Software Testing and Analysis, 2024, pp. 1541–1553

    work page 2024

  21. [21]

    An abstract domain for certifying neural networks,

    G. Singh, T. Gehr, M. P ¨uschel, and M. Vechev, “An abstract domain for certifying neural networks,”Proceedings of the ACM on Programming Languages, vol. 3, no. POPL, pp. 1–30, 2019

    work page 2019

  22. [22]

    Ef- ficient neural network robustness certification with general activation functions,

    H. Zhang, T.-W. Weng, P.-Y . Chen, C.-J. Hsieh, and L. Daniel, “Ef- ficient neural network robustness certification with general activation functions,”Advances in neural information processing systems, vol. 31, 2018

    work page 2018

  23. [23]

    Ai2: Safety and robustness certification of neural networks with abstract interpretation,

    T. Gehr, M. Mirman, D. Drachsler-Cohen, P. Tsankov, S. Chaudhuri, and M. Vechev, “Ai2: Safety and robustness certification of neural networks with abstract interpretation,” in2018 IEEE symposium on security and privacy (SP). IEEE, 2018, pp. 3–18

    work page 2018

  24. [24]

    Example guided synthesis of linear approxi- mations for neural network verification,

    B. Paulsen and C. Wang, “Example guided synthesis of linear approxi- mations for neural network verification,” inInternational Conference on Computer Aided Verification. Springer, 2022, pp. 149–170

    work page 2022

  25. [25]

    Formal security analysis of neural networks using symbolic intervals,

    S. Wang, K. Pei, J. Whitehouse, J. Yang, and S. Jana, “Formal security analysis of neural networks using symbolic intervals,” in27th USENIX Security Symposium (USENIX Security 18), 2018, pp. 1599–1614

    work page 2018

  26. [26]

    Improving neural network verification through spurious region guided refinement,

    P. Yang, R. Li, J. Li, C. Huang, J. Wang, J. Sun, B. Xue, and L. Zhang, “Improving neural network verification through spurious region guided refinement,” inTACAS 2021, ser. Lecture Notes in Computer Science, vol. 12651. Springer, 2021, pp. 389–408

    work page 2021

  27. [27]

    Branch and bound for piecewise linear neural network verification,

    R. Bunel, P. Mudigonda, I. Turkaslan, P. Torr, J. Lu, and P. Kohli, “Branch and bound for piecewise linear neural network verification,” Journal of Machine Learning Research, vol. 21, no. 2020, 2020

    work page 2020

  28. [28]

    The marabou framework for verification and analysis of deep neural networks,

    G. Katz, D. A. Huang, D. Ibeling, K. Julian, C. Lazarus, R. Lim, P. Shah, S. Thakoor, H. Wu, A. Zelji ´cet al., “The marabou framework for verification and analysis of deep neural networks,” inComputer Aided Verification: 31st International Conference, CAV 2019, New York City, NY, USA, July 15-18, 2019, Proceedings, Part I 31. Springer, 2019, pp. 443–452

    work page 2019

  29. [29]

    Reluplex: An efficient smt solver for verifying deep neural networks,

    G. Katz, C. Barrett, D. L. Dill, K. Julian, and M. J. Kochenderfer, “Reluplex: An efficient smt solver for verifying deep neural networks,” inComputer Aided Verification: 29th International Conference, CAV 2017, Heidelberg, Germany, July 24-28, 2017, Proceedings, Part I 30. Springer, 2017, pp. 97–117

    work page 2017

  30. [30]

    Efficient formal safety analysis of neural networks,

    S. Wang, K. Pei, J. Whitehouse, J. Yang, and S. Jana, “Efficient formal safety analysis of neural networks,”Advances in neural information processing systems, vol. 31, 2018

    work page 2018

  31. [31]

    Affine arithmetic: concepts and applications,

    L. H. De Figueiredo and J. Stolfi, “Affine arithmetic: concepts and applications,”Numerical algorithms, vol. 37, pp. 147–158, 2004

    work page 2004

  32. [32]

    R. E. Moore, R. B. Kearfott, and M. J. Cloud,Introduction to interval analysis. SIAM, 2009

    work page 2009

  33. [33]

    Fairness testing: A comprehensive survey and analysis of trends,

    Z. Chen, J. M. Zhang, M. Hort, M. Harman, and F. Sarro, “Fairness testing: A comprehensive survey and analysis of trends,”ACM Trans- actions on Software Engineering and Methodology, vol. 33, no. 5, pp. 1–59, 2024

    work page 2024

  34. [34]

    Verifying individual fairness in machine learning models,

    P. G. John, D. Vijaykeerthy, and D. Saha, “Verifying individual fairness in machine learning models,” inConference on Uncertainty in Artificial Intelligence. PMLR, 2020, pp. 749–758

    work page 2020

  35. [35]

    Automatic perturbation analysis for scalable certified robustness and beyond,

    K. Xu, Z. Shi, H. Zhang, Y . Wang, K.-W. Chang, M. Huang, B. Kailkhura, X. Lin, and C.-J. Hsieh, “Automatic perturbation analysis for scalable certified robustness and beyond,”Advances in Neural Information Processing Systems, vol. 33, pp. 1129–1141, 2020

    work page 2020

  36. [36]

    Fast and effective robustness certification,

    G. Singh, T. Gehr, M. Mirman, M. P ¨uschel, and M. T. Vechev, “Fast and effective robustness certification,” inNeurIPS 2018, Montr ´eal, Canada, 2018, pp. 10 825–10 836

    work page 2018

  37. [37]

    S. P. Boyd and L. Vandenberghe,Convex optimization. Cambridge university press, 2004

    work page 2004

  38. [38]

    Gurobi Optimizer Reference Manual,

    Gurobi Optimization, LLC, “Gurobi Optimizer Reference Manual,”

    work page

  39. [39]

    Available: https://www.gurobi.com

    [Online]. Available: https://www.gurobi.com

    work page

  40. [40]

    Becker, R

    B. Becker and R. Kohavi, “Adult,” UCI Machine Learning Repository, 1996, DOI: https://doi.org/10.24432/C5XW20

    work page doi:10.24432/c5xw20 1996

  41. [41]

    Statlog (German Credit Data)

    H. Hofmann, “Statlog (German Credit Data),” UCI Machine Learning Repository, 1994, DOI: https://doi.org/10.24432/C5NC77

    work page doi:10.24432/c5nc77 1994

  42. [42]

    MANCUR OLSON

    S. Moro, P. Rita, and P. Cortez, “Bank Marketing,” UCI Machine Learning Repository, 2012, DOI: https://doi.org/10.24432/C5K306

    work page doi:10.24432/c5k306 2012

  43. [43]

    Machine bias,

    J. Angwin, J. Larson, S. Mattu, and L. Kirchner, “Machine bias,” 2016. [Online]. Available: https://www.propublica.org/article/ machine-bias-risk-assessments-in-criminal-sentencing

    work page 2016

  44. [44]

    Evaluating robustness of neural networks with mixed integer programming,

    V . Tjeng, K. Y . Xiao, and R. Tedrake, “Evaluating robustness of neural networks with mixed integer programming,” in7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. [Online]. Available: https://openreview.net/forum?id=HyGIdiRqtm

    work page 2019

  45. [45]

    Certifying and removing disparate impact,

    M. Feldman, S. A. Friedler, J. Moeller, C. Scheidegger, and S. Venkata- subramanian, “Certifying and removing disparate impact,” inproceed- ings of the 21th ACM SIGKDD international conference on knowledge discovery and data mining, 2015, pp. 259–268

    work page 2015

  46. [46]

    Equality of opportunity in supervised learning,

    M. Hardt, E. Price, and N. Srebro, “Equality of opportunity in supervised learning,”Advances in neural information processing systems, vol. 29, 2016

    work page 2016

  47. [47]

    Learning fair representations,

    R. Zemel, Y . Wu, K. Swersky, T. Pitassi, and C. Dwork, “Learning fair representations,” inInternational conference on machine learning. PMLR, 2013, pp. 325–333

    work page 2013

  48. [48]

    Fairness through awareness,

    C. Dwork, M. Hardt, T. Pitassi, O. Reingold, and R. Zemel, “Fairness through awareness,” inProceedings of the 3rd innovations in theoretical computer science conference, 2012, pp. 214–226

    work page 2012

  49. [49]

    Fairness testing through extreme value theory,

    V . Monjezi, A. Trivedi, V . Kreinovich, and S. Tizpaz-Niari, “Fairness testing through extreme value theory,” in2025 IEEE/ACM 47th Inter- national Conference on Software Engineering (ICSE). IEEE, 2025, pp. 1501–1513

    work page 2025

  50. [50]

    Deepgemini: verifying depen- dency fairness for deep neural network,

    X. Xie, F. Zhang, X. Hu, and L. Ma, “Deepgemini: verifying depen- dency fairness for deep neural network,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 12, 2023, pp. 15 251– 15 259

    work page 2023

  51. [51]

    Fairer: fairness as decision rationale alignment,

    T. Li, Q. Guo, A. Liu, M. Du, Z. Li, and Y . Liu, “Fairer: fairness as decision rationale alignment,” inInternational Conference on Machine Learning. PMLR, 2023, pp. 19 471–19 489

    work page 2023

  52. [52]

    Fairness guarantees under demographic shift,

    S. Giguere, B. Metevier, Y . Brun, B. C. Da Silva, P. S. Thomas, and S. Niekum, “Fairness guarantees under demographic shift,” inProceed- ings of the 10th International Conference on Learning Representations (ICLR), 2022

    work page 2022

  53. [53]

    Autoric: Automated neural network repairing based on constrained optimization,

    X. Sun, W. Liu, S. Wang, T. Chen, Y . Tao, and X. Mao, “Autoric: Automated neural network repairing based on constrained optimization,” ACM Transactions on Software Engineering and Methodology, vol. 34, no. 2, pp. 1–29, 2025

    work page 2025

  54. [54]

    Nn repair: constraint-based repair of neural network classifiers,

    M. Usman, D. Gopinath, Y . Sun, Y . Noller, and C. S. P ˘as˘areanu, “Nn repair: constraint-based repair of neural network classifiers,” in Computer Aided Verification: 33rd International Conference, CAV 2021, Virtual Event, July 20–23, 2021, Proceedings, Part I 33. Springer, 2021, pp. 3–25

    work page 2021

  55. [55]

    Repairing misclassifica- tions in neural networks using limited data,

    P. Henriksen, F. Leofante, and A. Lomuscio, “Repairing misclassifica- tions in neural networks using limited data,” inProceedings of the 37th ACM/SIGAPP Symposium on Applied Computing, 2022, pp. 1031–1038

    work page 2022

  56. [56]

    Arachne: Search-based repair of deep neural networks,

    J. Sohn, S. Kang, and S. Yoo, “Arachne: Search-based repair of deep neural networks,”ACM Transactions on Software Engineering and Methodology, vol. 32, no. 4, pp. 1–26, 2023

    work page 2023

  57. [57]

    Inter- pretability based neural network repair,

    Z. Chen, J. Zhou, Y . Sun, J. Wang, Q. Xuan, and X. Yang, “Inter- pretability based neural network repair,” inProceedings of the 33rd ACM SIGSOFT International Symposium on Software Testing and Analysis, 2024, pp. 908–919

    work page 2024

  58. [58]

    Vere: Verification guided synthesis for repairing deep neural networks,

    J. Ma, P. Yang, J. Wang, Y . Sun, C.-C. Huang, and Z. Wang, “Vere: Verification guided synthesis for repairing deep neural networks,” in Proceedings of the 46th IEEE/ACM International Conference on Soft- ware Engineering, 2024, pp. 1–13

    work page 2024

  59. [59]

    Provable repair of deep neural networks,

    M. Sotoudeh and A. V . Thakur, “Provable repair of deep neural networks,” inProceedings of the 42nd ACM SIGPLAN International Conference on Programming Language Design and Implementation, 2021, pp. 588–603

    work page 2021

  60. [60]

    Sound and complete neural network repair with min- imality and locality guarantees,

    F. Fu and W. Li, “Sound and complete neural network repair with min- imality and locality guarantees,” inThe Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29,

    work page 2022

  61. [61]

    OpenReview.net, 2022

    work page 2022

  62. [62]

    Architecture-preserving provable repair of deep neural networks,

    Z. Tao, S. Nawas, J. Mitchell, and A. V . Thakur, “Architecture-preserving provable repair of deep neural networks,”Proceedings of the ACM on Programming Languages, vol. 7, no. PLDI, pp. 443–467, 2023. 0 50 100 150 200 Epoch 0.2 0.4 0.6 0.8 1.0Metric Value Compas - Gender 0 50 100 150 200 Epoch 0.2 0.4 0.6 0.8 1.0Metric Value Compas - Race 0 50 100 150 20...

    work page 2023