RESFL: An Uncertainty-Aware Framework for Responsible Federated Learning by Balancing Privacy, Fairness and Utility

1. Problem Statement

Federated Learning (FL) enables decentralized model training without sharing raw data, addressing privacy concerns in critical domains like healthcare and autonomous driving. However, existing FL frameworks face a critical triangular trade-off between Privacy, Fairness, and Utility:

• Privacy vs. Fairness: Standard privacy mechanisms (e.g., Differential Privacy) often obscure sensitive attributes (e.g., race, gender, skin tone) required to detect and correct bias. This inadvertently worsens performance disparities across demographic groups.
• Uncertainty and Reliability: Real-world data contains environmental noise (e.g., fog, rain) and domain shifts. Unmodeled uncertainty disproportionately degrades performance for vulnerable subpopulations (e.g., darker-skinned pedestrians in low-light conditions), amplifying fairness gaps.
• Adversarial Risks: Federated deployments are vulnerable to malicious clients who can poison updates or manipulate confidence signals, further inflating group disparities and privacy leakage.

Current solutions typically optimize one objective at the expense of others (e.g., high privacy with low fairness, or high fairness with high privacy leakage). There is a lack of a unified framework that simultaneously addresses privacy preservation, group fairness, and robustness to uncertainty in a single pipeline.

2. Methodology: The RESFL Framework

The authors propose RESFL (Responsible Federated Learning), a domain-agnostic framework that integrates two complementary components to jointly optimize privacy and fairness without sacrificing utility.

A. Adversarial Privacy Disentanglement (Client-Side)

To prevent sensitive attribute leakage during training, RESFL employs an adversarial representation learning approach:

• Gradient Reversal Layer (GRL): A feature extractor maps input data to a latent representation. An adversarial classifier attempts to predict the sensitive attribute (e.g., skin tone) from this representation.
• Minimax Objective: The feature extractor is trained to maximize the adversarial classifier's error (via the GRL), effectively removing sensitive attribute signals from the shared representations.
• Information-Theoretic Guarantee: By minimizing the mutual information between the latent representation and the sensitive attribute, the framework drives attribute inference attacks toward chance levels, ensuring privacy without needing to transmit raw attributes.

B. Uncertainty-Guided Fairness-Aware Aggregation (Server-Side)

To address fairness and reliability, RESFL introduces an Evidential Neural Network (ENN) and a novel Uncertainty Fairness Metric (UFM):

• Evidential Output: Instead of standard softmax, clients use an evidential head to predict a concentration vector ($\alpha$) parameterizing a Dirichlet distribution. This allows for the closed-form computation of epistemic uncertainty (model ignorance) without expensive Monte Carlo sampling.
• Uncertainty Fairness Metric (UFM):
  • Clients compute the average total evidence ($\bar{\alpha}_0$) for each demographic group.
  • UFM quantifies the disparity in uncertainty across groups: $UFM = \frac{\max(1/\bar{\alpha}_{0,g}) - \min(1/\bar{\alpha}_{0,g})}{\text{mean}(1/\bar{\alpha}_{0,g})}$.
  • A lower UFM indicates equitable confidence across groups; a higher UFM indicates disparity.
• Adaptive Aggregation: The server aggregates client updates using weights derived from the reported UFM:
  $$\omega_i \propto \exp(-\beta \cdot UFM_i)$$
  Clients with lower inter-group uncertainty disparity (lower UFM) and higher confidence receive higher weights. This prioritizes updates that are both accurate and equitable.
• Robustness Gating: A deterministic check blocks clients with validation accuracy below a threshold from receiving high weights, mitigating the impact of poisoned or low-confidence updates.

C. Joint Optimization

Each client minimizes a composite loss function:
$$\mathcal{L}_{local} = \mathcal{L}_{task} + \lambda_{priv}\mathcal{L}_{adv} + \lambda_{fair}\mathcal{L}_{uncertainty}$$
where $\mathcal{L}_{adv}$ enforces privacy via the GRL, and $\mathcal{L}_{uncertainty}$ (derived from the evidential head) encourages calibrated uncertainty estimates that inform the UFM.

3. Key Contributions

1. Unified Framework: RESFL is the first framework to jointly optimize privacy, fairness, and utility in FL using a single end-to-end pipeline, overcoming the "privacy-fairness trade-off" that plagues existing methods.
2. Novel Metrics & Mechanisms:
   • Introduction of the Uncertainty Fairness Metric (UFM), a scale-invariant measure of inter-group epistemic disparity.
   • Integration of Evidential Neural Networks for efficient uncertainty quantification in object detection.
   • Use of Gradient Reversal Layers for privacy-preserving disentanglement of sensitive attributes.
3. Theoretical Guarantees: The paper provides theoretical bounds showing that controlling UFM tightens confidence-adjusted generalization bounds for different groups, mathematically linking uncertainty minimization to fairness improvement.
4. Domain Agnosticism: While validated on autonomous driving, the framework is designed to be applicable to tabular and text data, demonstrated via experiments on Adult and TweetEval datasets.

4. Experimental Results

The authors evaluated RESFL on FACET (real-world images with skin tone annotations) and CARLA (autonomous driving simulator with weather variations), comparing against baselines like FedAvg, FedAvg-DP, FairFed, and PUFFLE.

• Utility (Accuracy):
  • On FACET, RESFL achieved 0.6654 mAP, outperforming FedAvg (0.6378) and PUFFLE (0.4192), and approaching FairFed (0.7013) while maintaining superior privacy.
  • It maintained robust performance across all 10 Monk Skin Tone (MST) cohorts, whereas FedAvg showed significant drops for darker skin tones.
• Fairness:
  • RESFL reduced the Equality of Opportunity (EOP) gap by 17% relative to the FedAvg baseline.
  • It achieved the lowest Disparate Impact deviation ($|1-DI| = 0.2287$) among high-utility models.
• Privacy:
  • Membership Inference Attack (MIA) Success Rate: Reduced by 37% compared to FedAvg (0.2093 vs. 0.3341).
  • Attribute Inference Attack (AIA) Success Rate: Significantly lower than all baselines (0.1832 vs. 0.4431 for FedAvg), demonstrating effective sensitive attribute removal.
• Robustness to Adversarial Conditions:
  • Under severe weather (fog, rain, cloud), RESFL maintained superior mAP and fairness compared to baselines.
  • In 100% fog conditions, where other models collapsed (mAP < 0.10), RESFL retained 0.17 mAP and kept fairness/privacy metrics significantly lower than competitors.
• Label-Free Variants: The framework was shown to retain >90% of its fairness/privacy benefits even when explicit sensitive labels were unavailable, using latent feature clustering (AF-UFM).

5. Significance and Impact

• Safety-Critical Applications: RESFL is particularly significant for autonomous vehicles and healthcare, where biased models can lead to life-threatening errors for marginalized groups. By ensuring equitable reliability under adverse weather and distribution shifts, it directly addresses safety concerns.
• Regulatory Compliance: The framework offers a practical path for organizations to deploy FL models that comply with emerging AI regulations (e.g., EU AI Act) requiring both privacy protection and fairness, without needing to centralize sensitive data.
• Scalability: The method adds only constant computational overhead (linear scaling) and minimal communication cost (scalar UFM), making it deployable on resource-constrained edge devices.
• Paradigm Shift: RESFL moves beyond the traditional "privacy vs. utility" or "fairness vs. utility" dichotomy, demonstrating that uncertainty-awareness is the key to unlocking a "responsible" regime where all three objectives can be optimized simultaneously.