Adversarial training protects predictive performance under attack but does not determine how decisions are distributed across protected groups. We formulate binary classification as projected gradient descent (PGD) adversarial-risk minimization under a squared soft demographic-parity (DP) budget, optimized with projected multiplier and one-sided penalty updates. On Adult Income, using the binary sex field supplied by the benchmark for group auditing but excluding it from prediction, we test logistic regression and a one-hidden-layer multilayer perceptron (MLP) across three paired seeds, four PGD strengths, and five budgets. At ε=0.3, PGD training improves balanced accuracy over empirical-risk minimization (ERM) by 7.77 and 11.45 percentage points and removes 85.7% and 89.5% of attack-induced growth in hard demographic-parity difference (DPD). At the representative interior value τ=0.015, constrained training cuts DPD relative to PGD training by 52.2% and 49.2%, while remaining 3.18 and 8.34 balanced-accuracy points above ERM. Across all eight seed-averaged model-attack paths, the sampled budget path strictly orders DPD and accuracy and reverses mean signed equal-opportunity difference (EOpp). A rate theorem gives the exact true-positive-rate shift required for reversal, and the measured group rates satisfy it. Within this controlled feature-space study, robustness and parity admit useful joint operating points, but one fairness metric cannot replace a full group-rate audit.
Figures & tables
Figure 1: Schematic overview. A parity budget coordinates robust training with group-rate control. Curves, axis values, and group bars are illustrative, not measured results. Section 3 gives the exact updates; Section 5 reports the operating paths and EOpp reversal.
Figure 2: Robust training stabilizes performance; the parity budget lowers DPD. Large markers are three-seed means and faint traces are individual seeds. PGD training limits attack-induced balanced-accuracy loss and DPD growth but begins from nearly the same clean DPD. Fair PGD at τ=0.015 lowers DPD across the evaluated range and remains above ERM in robust balanced accuracy at ε=0.3 . Here ε=0 is clean evaluation and DPD=PPR0−PPR1 .
Evaluation
Model
Training
BA
DPD
EOpp
ASR
Clean
Logistic
ERM
76.45±12
18.95±69
12.21±479
–
PGD train
74.02±19
18.48±52
17.32±427
–
Fair PGD
69.14±22
9.13±44
−2.97±437
–
Small MLP
ERM
78.27±25
18.64±91
7.99±463
–
PGD train
77.53±39
18.30±75
8.74±486
–
Fair PGD
74.32±94
10.04±157
−8.70±538
–
Table 1: Exact operating points confirm a joint robust–fair region. Three-seed mean ± sample standard deviation. Balanced accuracy (BA) and attack success rate (ASR) are percent; DPD and signed EOpp are percentage points. Fair PGD uses τ=0.015 ; ASR is omitted for clean input.
Figure 3: The DP budget traces a sampled operating path and can reverse equal-opportunity direction. Large markers are three-seed means; faint marks and traces are individual seeds. Left: tighter budgets lower hard DPD with a corresponding robust-accuracy cost at ε=0.3 ; outlined points mark τ=0.015 as an interior illustration, not an optimum. Right: mean signed EOpp changes direction as τ tightens, while seed variation shows that no single crossing budget applies across models and runs. Lines connect tested budgets only; “frontier” denotes this sampled path.
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Setting
Executed value
Dataset
OpenML Adult id 1590; complete cases; 45,222 records
Group field
sex ; Group 0 = Male, Group 1 = Female; excluded from predictor
Splits
Label-stratified 60/20/20 for seeds 0, 1, 2
Preprocessing
Six standardized continuous fields; seven one-hot categorical fields; training-fit only
Input dimension
102: 6 continuous and 96 one-hot coordinates
Models
Logistic regression; one-hidden-layer MLP with 64 ReLU units
Appendix
Table 3: Complete executed protocol. Training and evaluation PGD differ in initialization, step size, and range clipping.
Training
ε
BA
DPD
EOpp
ASR
ERM
0.0
76.45±0.12
18.95±0.69
12.21±4.79
–
ERM
0.1
73.08±0.12
20.05±0.84
11.80±4.66
3.68±0.10
ERM
0.2
68.82±0.42
22.04±0.85
14.09±5.64
8.56±0.51
ERM
0.3
64.12±0.48
23.85±0.79
13.98±4.78
14.33±0.68
PGD train
0.0
74.02±0.19
18.48±0.52
17.32±4.27
–
PGD train
0.1
73.28±0.11
18.62±0.54
16.94±4.40
0.77±0.16
Appendix
Table 4: Complete central PGD evaluation. Three-seed mean ± sample standard deviation. BA and ASR are percent; DPD and signed EOpp are percentage points. Fair PGD uses τ=0.015 .
Training
ε=0
0.1
0.2
0.3
ERM
0.9040±0.0042
0.8643±0.0032
0.8156±0.0016
0.7580±0.0017
PGD train
0.8890±0.0047
0.8785±0.0044
0.8680±0.0042
0.8568±0.0037
Fair PGD
0.8855±0.0059
0.8751±0.0052
0.8650±0.0044
0.8544±0.0037
Static aug.
0.8282
0.6486
0.4358
0.2459
Appendix
Table 5: Complete AUC record for the central evaluation. Central methods report three-seed mean ± sample standard deviation. Static augmentation is aggregate-only.
Model
Training
Seed
BA
DPD
EOpp
ASR
Logistic regression
ERM
0
64.65
23.09
+10.50
13.64
1
63.99
24.67
+19.43
14.36
2
63.71
23.80
+12.02
15.00
PGD train
0
72.13
19.28
+14.21
2.08
1
71.53
19.74
+21.11
2.56
2
71.99
18.52
+12.87
2.48
Appendix
Table 6: Paired seed-level values at PGD ε=0.3 . BA and ASR are percent; DPD and signed EOpp are percentage points.
τ
ε=0
0.1
0.2
0.3
0.025
72.68±0.20
72.12±0.07
71.41±0.06
70.67±0.10
0.020
71.29±0.20
70.59±0.32
69.73±0.31
69.19±0.55
0.015
69.14±0.22
68.55±0.27
67.94±0.29
67.30±0.18
0.010
67.31±0.32
66.58±0.51
65.94±0.73
65.33±0.81
0.005
61.12±1.62
60.44±1.44
59.76±1.36
58.98±1.33
Appendix
Table 7: Complete fairness-budget grid for logistic regression. Three-seed mean ± sample standard deviation under PGD evaluation.
τ
ε=0
0.1
0.2
0.3
0.025
76.39±0.24
75.54±0.17
74.49±0.14
73.07±0.25
0.020
75.47±0.32
74.55±0.37
73.33±0.34
72.11±0.34
0.015
74.32±0.94
73.53±1.04
72.26±0.91
71.09±0.71
0.010
72.95±0.42
72.00±0.40
71.00±0.50
69.89±0.64
0.005
69.93±0.42
69.05±0.42
68.15±0.32
66.95±0.38
Appendix
Table 8: Complete fairness-budget grid for small MLP. Three-seed mean ± sample standard deviation under PGD evaluation.
Figure 4: Complete fairness-budget paths. Three-seed means for Fair PGD across every budget and PGD evaluation strength. Columns show BA, hard DPD, and signed EOpp; rows separate logistic regression and the small MLP. Lines connect sampled budgets and do not assert global Pareto optimality or a hard-DPD guarantee.
τ
ε=0
0.1
0.2
0.3
0.025
0.8887±0.0050
0.8785±0.0046
0.8684±0.0042
0.8577±0.0039
0.020
0.8876±0.0053
0.8775±0.0049
0.8676±0.0044
0.8571±0.0040
0.015
0.8855±0.0059
0.8751±0.0052
0.8650±0.0044
0.8544±0.0037
0.010
0.8816±0.0066
0.8709±0.0059
0.8605±0.0052
0.8496±0.0048
0.005
0.8719±0.0087
0.8612±0.0080
0.8509±0.0072
0.8401±0.0066
Appendix
Table 9: AUC along the complete fairness-budget path. Three-seed mean ± sample standard deviation.
Training
ε
PPR 0
PPR 1
TPR 0
TPR 1
ERM
0.0
26.20±0.38
7.26±0.36
61.62±0.92
49.41±3.88
ERM
0.1
27.79±0.40
7.74±0.45
57.71±0.74
45.91±3.93
ERM
0.2
30.35±0.35
8.31±0.50
53.55±1.12
39.46±4.74
ERM
0.3
33.71±0.43
9.86±0.42
49.22±0.80
35.23±4.17
PGD train
0.0
24.86±0.44
6.38±0.44
57.55±0.60
40.22±3.68
PGD train
0.1
25.11±0.45
6.49±0.46
56.58±0.63
39.65±3.91
Appendix
Table 10: Complete group-level metrics for logistic regression. Three-seed mean ± sample standard deviation. Group 0 is Male and Group 1 is Female.
Training
ε
PPR 0
PPR 1
TPR 0
TPR 1
ERM
0.0
27.56±0.22
8.92±0.79
65.16±0.95
57.17±3.69
ERM
0.1
30.48±0.29
9.89±0.90
60.68±0.86
52.60±4.37
ERM
0.2
34.45±0.44
11.31±1.16
55.92±0.88
48.42±5.19
ERM
0.3
38.07±0.36
13.85±1.43
50.25±0.63
43.83±4.93
PGD train
0.0
26.74±0.12
8.44±0.70
63.45±1.01
54.72±3.94
PGD train
0.1
26.79±0.21
8.51±0.64
62.16±0.73
53.26±3.50
Appendix
Table 11: Complete group-level metrics for small MLP. Three-seed mean ± sample standard deviation. Group 0 is Male and Group 1 is Female.
Figure 5: Group-rate paths under the fairness budget. Three-seed mean PPR and TPR at PGD ε=0.3 as τ tightens. Group 0 is Male and Group 1 is Female. The figure exposes the rate redistribution behind the aggregate DPD and EOpp values.
Model
ε
Attack
BA
DPD
ASR
Logistic regression
0.1
FGSM
73.08±0.12
20.05±0.84
3.68±0.10
0.1
PGD
73.08±0.12
20.05±0.84
3.68±0.10
0.2
FGSM
68.82±0.42
22.04±0.85
8.56±0.51
0.2
PGD
68.82±0.42
22.04±0.85
8.56±0.51
0.3
FGSM
64.12±0.48
23.85±0.79
14.33±0.68
0.3
PGD
64.12±0.48
23.85±0.79
14.33±0.68
Appendix
Table 12: FGSM and PGD evaluation of ERM models. Three-seed mean ± sample standard deviation. BA and ASR are percent; DPD is percentage points.
Figure 6: FGSM and PGD evaluation of ERM models. The panels compare BA, DPD, and ASR over the tested strengths for both model classes. PGD remains the principal iterative white-box evaluation; close agreement does not constitute certification or prove attack optimality.
Model
Attack
ε=0
0.1
0.2
0.3
Logistic regression
FGSM
0.9040±0.0042
0.8643±0.0032
0.8156±0.0016
0.7580±0.0017
PGD
0.9040±0.0042
0.8643±0.0032
0.8156±0.0016
0.7580±0.0017
Small MLP
FGSM
0.9111±0.0034
0.8645±0.0035
0.8046±0.0028
0.7321±0.0022
PGD
0.9111±0.0034
0.8645±0.0035
0.8044±0.0029
0.7314±0.0023
Appendix
Table 13: AUC under FGSM and PGD evaluation of ERM models. Three-seed mean ± sample standard deviation.
Model
ε
BA static
BA online
ASR static
ASR online
Logistic regression
0.0
69.56
74.02±0.19
–
–
0.1
54.98
73.28±0.11
15.44
0.77±0.16
0.2
40.38
72.54±0.29
32.20
1.59±0.25
0.3
29.58
71.88±0.31
46.80
2.37±0.26
Small MLP
0.0
78.61
77.53±0.39
–
–
0.1
34.12
76.61±0.26
45.96
0.84±0.09
Appendix
Table 14: Static adversarial augmentation is a negative control. Static rows are aggregate means; seed-level evaluation values were not retained. Fairness columns are excluded because the retained static fairness views did not pass the cross-file audit.
Artifact
File
SHA-256 prefix
Main source
main.tex
b1b62a6ab2
Appendix source
appendix_content.tex
e15441ace7
Bibliography
references.bib
11e8e45556
Appendix generator
generate_appendix_assets.py
2f4f0393f4
ERM PGD results
pgd_combined.csv
816adfa334
PGD-training results
pgd_train_combined.csv
e8ac869198
Appendix
Table 15: Original analysis artifact record. File names and SHA-256 prefixes identify artifacts recorded during the analysis. The source-file hashes refer to that analysis snapshot, not to this revised manuscript. The table is a provenance record, not a list of files supplied with the typesetting source.
Algorithmic fairness studies how automated decisions distribute errors, opportunities, and harms across groups, and how to measure and reduce disparities.
Machine learning models trained on sensitive data can inadvertently leak population-level information about their training distributions -- a threat known as distribution inference attack (DIA). An adversary with black-box access can infer sensitive demographic properties, such as subgroup proportions, without observing any training data directly. While defenses such as differential privacy and property unlearning have been proposed, the link between fairness constraints and distributional leakage remains unexplored. We propose Fair Fine-tuning (FFt): a trained model is fine-tuned on samples from the complementary distribution under an Equalized Odds (EO) constraint. We provide a complete theoretical characterization, proving the tight bound Adv(A,Mf)≤ΔEO⋅W, where W quantifies how distinguishable the two training distributions are by their sensitive-attribute composition. We also establish a necessary condition for FFt to reduce adversarial advantage and prove tightness of the bound. We evaluate across six datasets spanning tabular (ACS Income, COMPAS, German Credit), image (UTKFaces), and NLP (Bias in Bios) modalities. Rehearsal-based FFt consistently reduces the adversarial accuracy gap below the detection threshold τ!=!0.1 across all settings; on ACS Income, the gap falls from ∼!15 to under 4. Our work provides the first formal bound connecting a model's measured EO disparity directly to its adversarial advantage in the DIA game, opening a new avenue for unified fairness-and-privacy defenses.
We focus on regression settings where the fairness goal is to equalize average predictions across values of a sensitive attribute, a criterion known as mean demographic parity. Directly optimizing this criterion is difficult because it depends on a conditional mean that is unknown and changes during training. Common dependence penalties and adversarial methods do not estimate this conditional mean; instead, they push predictions toward full independence. This stronger constraint can reduce accuracy even when average predictions are already equal. Existing conditional-mean methods are limited to linear predictors or low-dimensional sensitive attributes. We enable direct optimization of mean demographic parity using DPVar, a fairness measure defined as the variance of the conditional mean prediction. Because the conditional mean must be estimated as the predictor changes, optimizing DPVar leads to a functional bilevel problem. We develop two solvers: FBO, which uses a closed-form hypergradient, and an iterative-differentiation (ITD) solver that differentiates through updates of the conditional-mean estimator. Unlike previous conditional-mean methods, our approach applies to nonlinear predictors and high-dimensional continuous sensitive attributes. Across a semi-synthetic benchmark built from 21 tabular regression datasets and Communities & Crime data, FBO and ITD recover competitive or better accuracy-DPVar trade-offs than existing methods.
Differential privacy (DP) is increasingly deployed to limit membership inference risk in machine-learning systems. Prior work has shown that DP-SGD can widen accuracy disparities across demographic groups, but this framing treats fairness as a purely outcome-side concern. We argue that privacy cost, the information leakage borne by each group, is itself a form of harm, and adopt a compensatory-fairness framework in which a group that involuntarily bears greater privacy exposure is owed proportionally greater benefit from the system. From this principle we derive the Privacy-Cost Equity Ratio (PCER), a group fairness metric defined as a group's positive prediction rate normalized by its per-group overfitting gap. By a standard membership inference bound, this overfitting gap upper-bounds each group's vulnerability to inference attacks, making PCER a conservative measure of benefit relative to exposure. PCER needs only per-group train and test accuracy (no shadow models), making it a practical post-hoc audit tool. We evaluate PCER alongside standard fairness metrics across six benchmark--attribute combinations spanning tabular and NLP domains, under DP-SGD at a range of privacy budgets, and validate the overfitting-gap proxy against a direct threshold membership-inference attack. The results reveal patterns that outcome-based metrics miss. On COMPAS, PCER uncovers a persistent double disadvantage: the protected group bears both greater privacy exposure and worse predictive outcomes, something demographic parity gap masks entirely. Sensitivity analysis shows very strong privacy guarantees collapse both groups' overfitting to a numerical floor, rendering exposure-based audits uninformative in that regime. Together, these findings show that fairness audits of privacy-preserving systems must account for who bears the cost of protection, not only who benefits from its outcomes.