AI IN FINANCIAL SERVICES
The PGD Paradox in Credit Scoring: When Stronger Attacks Improve Accuracy but Worsen Fairness
Research into adversarial robustness within credit scoring reveals a paradox: stronger iterative attacks (PGD) paradoxically enhance accuracy compared to single-step attacks (FGSM), which is counter-intuitive to traditional beliefs. We observe that this "PGD Paradox" occurs consistently and generates systematic fairness violations in which privileged populations get a disproportionate benefit (mean benefit inequity=7.57%). We test robustness with realistic feature restrictions, and the results are reproducible, which demonstrates that findings are not the result of invalid perturbation.
Quantifying the Paradox & Its Cost
Stronger adversarial attacks (PGD) surprisingly boost model accuracy but exacerbate fairness issues for vulnerable groups. Our analysis provides crucial metrics for understanding these complex trade-offs in credit scoring.
0
Accuracy Enhancement with PGD
0
Average Benefit Inequity
0
Max RFS Gap (Worst Group)
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
The PGD Paradox: Stronger Attacks, Higher Accuracy
Contrary to traditional beliefs, our research uncovers a PGD Paradox in credit scoring: stronger iterative attacks (Projected Gradient Descent - PGD) surprisingly *enhance* model accuracy compared to weaker single-step attacks (Fast Gradient Sign Method - FGSM). This counter-intuitive phenomenon is consistent across diverse financial datasets, demonstrating an average accuracy enhancement of 15.08% under PGD attacks. This suggests that the 'strength' of an attack isn't solely about magnitude, but also about the quality and manifold alignment of the perturbations it generates, effectively acting as a regularization mechanism in tabular data.
Fairness Erosion: Disproportionate Benefits for Privileged Groups
While PGD attacks improve aggregate accuracy, they systematically create severe fairness violations. Privileged demographic groups consistently experience disproportionately higher benefits, leading to a 'rich-get-richer' effect. We measure this through Equalized Accuracy Disparity, revealing an average benefit inequity of 7.57%. In extreme cases like German Credit, this inequity reaches 15.09%, far exceeding typical regulatory concern thresholds. The Robustness Fairness Score (RFS) highlights these hidden inequities, showing an aggregate-to-worst-group gap of up to 23.0%, which would remain invisible with conventional aggregate metrics.
Towards Fairness-Aware Adversarial Evaluation
To address the PGD Paradox and its fairness implications, we propose a multi-tiered regulatory framework. This includes stratified reporting of accuracy and fairness metrics by demographic group, establishing clear fairness thresholds (e.g., RFS gap < 3% for minimal risk), and mandating multi-attack testing (FGSM and PGD) to uncover diverse weaknesses. Critical constraint validation ensures perturbations remain realistic. This framework aims to enforce equitable AI outcomes in high-stakes systems, transforming adversarial robustness from a purely technical concern to a core ethical responsibility.
15.08%
Average Accuracy Enhancement with PGD over FGSM
The PGD Paradox consistently improves accuracy on average across all datasets.
7.57%
Average Benefit Inequity from PGD Paradox
Stronger attacks generate systematic fairness violations, disproportionately benefiting privileged populations.
Regulatory Framework for AI Robustness
Tier 1 (Minimal Risk)
→
Tier 2 (Moderate Risk)
→
Tier 3 (Severe Risk)
| Dataset | FGSM L∞ | PGD L∞ | Ratio |
|---|---|---|---|
| Taiwan Credit | 0.300 | 0.180 | 0.60 |
| German Credit | 0.300 | 0.180 | 0.60 |
| South German Credit | 0.300 | 0.176 | 0.59 |
| HELOC | 0.300 | 0.179 | 0.60 |
Validating Realistic Perturbations
Our experiments rigorously confirmed that the PGD Paradox and its fairness implications are robust and not artifacts of invalid perturbations. Even with explicit domain-valid feature constraints applied (bounding values within the 5th-95th percentiles of training data), the results for accuracy, fairness gap, and benefit inequity were identical (p > 0.99) to unconstrained experiments. This demonstrates that PGD inherently tends to gravitate towards realistic feature combinations, unlike FGSM which often produces unrealistic values. This finding is crucial, as it validates our reported weaknesses as real-world adversarial situations, not artificial ones.
Calculate Your Potential AI Impact
Estimate the efficiency gains and cost savings your enterprise could achieve by implementing robust, fairness-aware AI solutions, based on industry averages.
Annual Savings
$0
Hours Reclaimed Annually
0
Your Path to Robust & Fair AI
Implementing fairness-aware adversarial robustness is a strategic journey. Here’s a typical roadmap for enterprise adoption:
Phase 1: Initial Adversarial Risk Assessment
Conduct comprehensive audits of existing AI systems using diverse attack methods (FGSM, PGD) and fairness metrics like Equalized Accuracy Disparity and RFS. Identify areas of high benefit inequity or significant RFS gaps.
Phase 2: Develop Fairness-Aware Robustness Strategies
Implement fairness-aware adversarial training techniques. Establish clear fairness thresholds (e.g., <7% RFS gap) and develop corrective action plans for models exceeding these thresholds. Prioritize interventions for severe risk cases (e.g., >10% benefit inequity).
Phase 3: Continuous Oversight & Adaptive Defense
Set up a continuous monitoring framework with stratified reporting for all demographic groups. Regularly review model performance, adversarial robustness, and fairness metrics. Adapt defense strategies to new attack types and evolving data distributions, ensuring long-term equitable outcomes.
Ready to Build Fair & Robust AI Systems?
Navigate the complexities of adversarial robustness and algorithmic fairness with confidence. Our experts are ready to help you implement a secure and equitable AI strategy.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat