PaperlayerView exact revision on arXiv

Source-linked AI summary

FairNeuron: Improving Deep Neural Network Fairness with Adversary Games on Selective Neurons

Xuanqi Gao, Juan Zhai, Shiqing Ma, Chao Shen, Yufei Chen, Qian Wang

arXiv:2204.02567v2cs.LGcs.AIcs.SE

TL;DR

Fairness-aware DNN decision systems must balance utility and fairness, although optimizing accuracy and fairness can conflict. FairNeuron locates conflict paths and selectively retrains affected samples without an adversary model; across three datasets, it improved fairness while maintaining stable utility.

  • Problem

    Fairness-aware DNN decision systems need high utility and fairness, but optimizing accuracy and fairness objectives can conflict.

    Source: §1 · p. 1, §2.3 · p. 4

  • Method

    FairNeuron identifies conflict paths, clusters samples that trigger sensitive-feature selection, and selectively retrains affected samples by muting neurons outside those paths.

    Source: §1 · p. 2

  • Results

    Across three datasets, FairNeuron maintained stable utility while improving fairness; average DP improvement was 57.65%, 20% higher than Ethical Adversaries.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Lightweight path analysis and dropout make FairNeuron a practical alternative that uses one-fifth the training time of Ethical Adversaries on average.

    Source: §1 · p. 2

    Adversarial repair alternatives can be difficult to train, with no guarantee that their adversary networks will converge.

    Source: §2.3 · p. 3

Abstract

from arXiv · show

With Deep Neural Network (DNN) being integrated into a growing number of critical systems with far-reaching impacts on society, there are increasing concerns on their ethical performance, such as fairness. Unfortunately, model fairness and accuracy in many cases are contradictory goals to optimize. To solve this issue, there has been a number of work trying to improve model fairness by using an adversarial game in model level. This approach introduces an adversary that evaluates the fairness of a model besides its prediction accuracy on the main task, and performs joint-optimization to achieve a balanced result. In this paper, we noticed that when performing backward propagation based training, such contradictory phenomenon has shown on individual neuron level. Based on this observation, we propose FairNeuron, a DNN model automatic repairing tool, to mitigate fairness concerns and balance the accuracy-fairness trade-off without introducing another model. It works on detecting neurons with contradictory optimization directions from accuracy and fairness training goals, and achieving a trade-off by selective dropout. Comparing with state-of-the-art methods, our approach is lightweight, making it scalable and more efficient. Our evaluation on 3 datasets shows that FairNeuron can effectively improve all models' fairness while maintaining a stable utility.

1 INTRODUCTION

FairNeuron repairs DNN fairness problems by identifying conflict paths and selectively retraining affected samples, without adding an adversary model. The paper reports improved fairness with stable utility and lower training time than Ethical Adversaries.

1 INTRODUCTION

  • Fairness repair matters for decision systems such as loan or interview screening, and the paper considers demographic parity, demographic parity rate, and equal opportunity.The introduction motivates fairness because biased decisions can affect specific social subgroups.Source: §1 · p. 1
  • Adversarial fairness-fixing approaches can be time-consuming and challenging, and their training protocols are heavyweight.The introduction also notes practical problems such as mode collapse.Source: §1 · pp. 1–2
  • FairNeuron identifies conflict paths, clusters samples that trigger sensitive-feature selection, then selectively retrains by muting neurons outside those paths for affected samples.Conflict paths contain neurons that select sensitive attributes rather than distinguishable features.Source: §1 · p. 2
  • FairNeuron achieved twice the fairness performance and used one-fifth the training time of Ethical Adversaries on average, using lightweight path analysis and dropout.The approach is implemented as a self-contained toolkit.Source: §1 · p. 2
  • Across three datasets, FairNeuron maintained stable utility; average DP improvement was 57.65%, 20% higher than Ethical Adversaries.The authors also report that FairNeuron improves fairness while maintaining stable utility.Source: §1 · p. 2
  • The proposed framework avoids training an adversary model and modifying the model’s training protocol or architecture.Its analysis and repair procedures are described as lightweight.Source: §1 · p. 2
  • The evaluation covers three datasets; configurable parameters may not transfer to substantially different model sizes, and the path-activation assumption may not hold for some models.The authors describe these as threats to validity.Source: §1 · p. 2

2 BACKGROUND AND MOTIVATION

The section reviews fairness notions and existing approaches, then motivates FairNeuron by locating accuracy–fairness conflicts in individual neurons and proposing a selective, adversary-free repair strategy.

2.1 Fairness

  • Fairness may concern equal treatment of individuals or of subpopulations defined by sensitive attributes, and the study selects DP, DPR, and EO as representative notions.Prior work reports strong correlations among fairness definitions, motivating selection of a few representative measures.Source: §2.1 · pp. 2–3
  • DP measures the difference in positive-class prediction probabilities across groups, whereas DPR expresses group prediction equality as a ratio.The ideal values are DP = 0 and DPR = 1, respectively.Source: §2.1 · pp. 2–3
  • EO compares positive-class prediction probabilities across groups conditional on the true label being positive, with zero indicating ideal fairness.In practice, the paper describes fairness thresholds for DP, DPR, and EO rather than requiring their ideal values.Source: §2.1 · p. 3

2.2 Improving DNN Fairness

  • FAD uses gradient reversal with predictor and adversary networks, while Ethical Adversaries iteratively train a reader and feeder to assess and address bias.FAD also uses post-training threshold adjustment based on ROC analysis.Source: §2.2 · p. 3
  • Other fairness fixes include pre-processing by reweighting group samples and post-processing by changing outcomes within an uncertain decision-boundary band.Reject option classification favors unprivileged groups and disfavors privileged groups within that band.Source: §2.2 · p. 3

2.3 Motivation and Basic Idea

  • Existing adversarial approaches can encounter mode collapse, vanishing gradients, or nonconvergence, and their games lack a guaranteed equilibrium when accuracy and fairness do not always conflict.The paper notes empirical cases where FAD worsens fairness on particular metrics and datasets.Source: §2.3 · pp. 3–4
  • Accuracy and fairness objectives can point in contradictory directions during training, while neurons represent feature combinations whose weights reflect feature importance.Accuracy-only optimization may favor directions that reduce fairness, including reliance on sensitive features.Source: §2.3 · p. 4
  • FairNeuron monitors neurons with conflicting accuracy and fairness optimization, identifies samples behind the conflicts, and seeks a balanced direction without adding an adversary.The proposed process is intended to simplify training and be more lightweight than existing solutions.Source: §2.3 · p. 4

3 DESIGN

FairNeuron repairs a biased model by locating paths associated with fairness conflicts, grouping the corresponding samples, and selectively retraining without changing the model structure. Its pipeline leaves benign samples on ordinary training and applies dropout training to biased samples.

3.1 Overview

  • FairNeuron takes a biased model and training data through path detection, sample clustering, and selective retraining to produce a fixed model.Benign samples retain ordinary training, while biased samples receive a different training strategy.Source: §3.1 · p. 4

3.2 Neural Network Slicing

  • Neural network slicing estimates neuron behavior from average activations and per-sample differences, then backtracks contributions through neurons and synapses to identify paths.The procedure starts at the output and uses a threshold based on subsequent-neuron activation and hyperparameter γ to select key synapses.Source: §3.2 · p. 5

3.3 Sample Clustering

  • Sample clustering marks paths with activation frequency below θ times the maximum frequency as biased, then groups their corresponding samples as biased samples.The method uses recorded path-to-sample relationships to divide the training data into biased and ordinary groups.Source: §3.3 · p. 6

3.4 Selective Training

  • Selective training deactivates dropout for ordinary samples and activates it for biased samples, encouraging the detected neurons to learn less biased features without changing model structure.For biased samples, the method seeks to encourage learning unbiased features rather than biased ones.Source: §3.4 · p. 6

4.1 Experiment Setup

The experiments use three fairness datasets and fully connected neural networks, with hyperparameters tuned on a subset of training data. They compare FairNeuron with in-processing, pre-processing, and post-processing baselines while examining effectiveness, efficiency, parameter effects, and image-dataset performance.

4.1.1 Hardware and software.

  • Experiments ran on a server with 32 CPU cores, 256 GB memory, and one NVIDIA TITAN V GPU under Ubuntu 16.04.Source: §4.1.1 · p. 7

4.1.2 Datasets.

  • The evaluation uses UCI Adult Census, COMPAS, and German Credit, with gender sensitive for Adult and German Credit and race sensitive for COMPAS.Adult has 32,561 instances and 9 features; German Credit has 600 records and 20 attributes.Source: §4.1.2 · p. 7

4.1.3 Models.

  • Each dataset uses a fully connected three-hidden-layer network; layers have 32 neurons for COMPAS and German Credit and 128 for Adult Census.Data are split into training, validation, and test sets at a 7:1:2 ratio.Source: §4.1.3 · p. 7

4.1.4 Hyperparameter tuning.

  • Hyperparameters θ and γ are selected by parallel grid search over specified ranges, using a small subset drawn from the training set.The search samples θ over [10^-4, 1] and γ over [0.5, 1], and uses Ray Tune.Source: §4.1.4 · p. 7

4.1.5 Metrics and baseline methods.

  • The baselines include FAD and Ethical Adversaries, reweighing, and Reject Option Classification; the experiments ask about effectiveness, efficiency, parameter effects, and image datasets.Source: §4.1.5 · p. 8

4.2 Effectiveness of FairNeuron

FairNeuron improves fairness across the evaluated datasets while largely preserving utility. Its slicing, clustering, and selective-training components are each evaluated, with selective training outperforming ordinary and pure-dropout training on the reported metrics.

4.2 Effectiveness of FairNeuron

  • Average utility is highest among models with fairness constraints, exceeding ROC by 27.9%, Reweighing by 17.5%, Ethical Adversaries by 3.85%, and FAD by 27.22%.FairNeuron also surpasses the naive model on German Credit and COMPAS; the reported average accuracy changes are -0.83%, 1.36%, and 28.66%.Source: §4.2 · p. 8

4.2.1 Effectiveness of neuron network slicing.

  • The slicing example uses a threshold of 1.41 from a maximum path-activation statistic of 47 and θ = 0.03; detected abnormal paths cluster near 1.Their corresponding sample proportion is low; Figure 4 encodes path-activation density, cumulative sample ratio, and the threshold.Source: §4.2.1 · p. 8, Figure 4 · p. 8

4.2.2 Effectiveness of sample clustering.

  • With FairNeuron’s sample clustering, average accuracy improves by 6.68%, alongside reported fairness improvements of 96.19% in DP, 97.67% in DPR, and 38.95% in EO.The comparison is against random clustering.Source: §4.2.2 · p. 8, Table 2 · p. 9

4.2.3 Effectiveness of selective training.

  • Selective training surpasses ordinary training by 38.96%, 98.22%, 97.43%, and 91.50%, and pure dropout by 22.27%, 96.19%, 97.55%, and 38.95%, in Acc, DP, DPR, and EO.The comparison evaluates ordinary, pure-dropout, and selective training.Source: §4.2.3 · pp. 8–9, Table 3 · p. 9

4.2 Effectiveness of FairNeuron

  • FairNeuron improves naive-model DPR by 98.47%, 157.23%, and 3895.23% on Census, Credit, and COMPAS, respectively, but Credit’s EO and DP results are unsatisfactory.It achieves the best fairness performance on Census and COMPAS; the authors attribute Credit’s weaker results to incomplete slicing on its 600 instances.Source: §4.2 · p. 8, Table 4 · p. 9

4.3 Efficiency of FairNeuron

FairNeuron’s repair procedure is faster than Ethical Adversaries on average, though its overhead relative to ordinary training varies by dataset. Hyperparameter tuning accounts for a larger share of runtime, especially on smaller datasets.

4.3 Efficiency of FairNeuron

  • FairNeuron uses 55.49% of Ethical Adversaries’ time on average and 5.39 times the time of ordinary training.The reported time comparison uses Ethical Adversaries’ per-iteration time over 50 iterations and FairNeuron’s per-trial time.Source: Table 5 · p. 9, §4.3 · p. 9
  • FairNeuron is reported as more efficient than Ethical Adversaries, with an average speedup of 180%.Hyperparameter tuning takes a larger share of its total runtime than ordinary training; the authors report less than twice ordinary-training time on Census, but relatively long runtime on German Credit.Source: §4.3 · p. 9, Table 6 · p. 9

4.4 Effects of Configurable Hyperparameters

FairNeuron’s configurable θ and γ affect neuron-network slicing and path complexity. The reported performance is generally not sensitive to their assignments, except for EO.

4.4 Effects of Configurable Hyperparameters

  • γ is the neuron-activation threshold: decreasing it includes more neurons and synapses, producing more complex paths.Source: §4.4 · p. 9
  • θ is the neuron-network-slicing threshold; lowering θ results in fewer paths classified as abnormal.The tested intervals were [10−4, 1] for θ and [0.5, 1] for γ.Source: §4.4 · p. 9
  • FairNeuron performs better than the naive model and Ethical Adversaries across the tested hyperparameter assignments, with sensitivity excepted for EO.Increasing EO’s weight in the hyperparameter-tuning loss function can constrain EO fluctuations.Source: §4.4 · pp. 9–10

4.5 Performance on Image Datasets

FairNeuron is evaluated on MNIST and CIFAR-10 image models using class-wise fairness metrics, and the reported results show improved fairness.

4.5 Performance on Image Datasets

  • On MNIST and CIFAR-10, FairNeuron improves fairness by 20% for MCD and 80% for CV.The comparison uses a four-layer fully connected network on MNIST and ResNet-18 on CIFAR-10, against naive models and random dropout.Source: §4.5 · p. 10

5 RELATED WORK

Related work surveys neural-network slicing, fairness definitions and testing, and approaches for repairing biased models.

5 RELATED WORK

  • Neural-network slicing adapts software path and dataflow analysis, including critical routing paths, execution-path profiling [56], and dynamic slicing [73].Related work also documents fairness concerns in higher education, employment [57], and re-offence judgment.Source: §5 · p. 10
  • Fairness notions include individual treatment of similar people [26] [46], equal treatment across groups [28] [35] [70], and per-group max-min fairness [45].These are presented as three broad categories.Source: §5 · p. 10
  • Fairness testing methods generate discriminatory test cases using causal analysis, random generation, input-space search, or symbolic execution with local explanations.THEMIS, AEQUITAS, ADF, and Symbolic Generation represent these approaches, respectively.Source: §5 · p. 10
  • Repair approaches are grouped into pre-processing of data, in-processing changes to training, and post-processing changes to model outputs.Examples include revising labels or attributes, adding fairness constraints or objectives, and changing predictive labels.Source: §5 · p. 10

6 CONCLUSION

FairNeuron combines path analysis and dropout training to repair fairness problems in DNNs. The evaluation reports improved effectiveness and efficiency, while CNN performance is constrained by analysis of only the last fully connected layer.

6 CONCLUSION

  • FairNeuron searches for biased instances through path analysis and mitigates fairness problems with dropout training.The conclusion presents the algorithm as a lightweight approach for improving DNN performance.Source: §6 · p. 10
  • The evaluation reports significantly better effectiveness and efficiency in fixing biased models.The conclusion gives no quantitative result in this passage.Source: §6 · p. 10
  • For CNNs, FairNeuron can be applied only to the last fully connected layer, and its performance is not ideal.The authors identify improving CNN support as future work.Source: §6 · p. 10

Related papers

Loading 2204.02567v2…