Federated Learning (FL) has gained prominence in machine learning applications across critical domains by enabling collaborative model training without centralized data aggregation. However, FL frameworks that protect privacy often sacrifice fairness and reliability. Differential privacy can reduce data leakage, but it may also obscure sensitive attributes needed for bias correction, thereby worsening performance gaps across demographic groups. This work studies the privacy-fairness trade-off in FL-based object detection and introduces RESFL, an integrated framework that jointly improves both objectives. RESFL combines adversarial privacy disentanglement with uncertainty-guided fairness-aware aggregation. The adversarial component uses a gradient reversal layer to suppress sensitive attribute information, reducing privacy risks while preserving fairness-relevant structure. The uncertainty-aware aggregation component uses an evidential neural network to adaptively weight client updates, prioritizing contributions with lower fairness disparities and higher confidence. This produces robust and equitable FL model updates. Experiments in high-stakes autonomous vehicle settings show that RESFL achieves high mAP on FACET and CARLA, reduces membership-inference attack success by 37%, reduces the equality-of-opportunity gap by 17% relative to the FedAvg baseline, and maintains stronger adversarial robustness. Although evaluated in autonomous driving, RESFL is domain-agnostic and can be applied to a broad range of application domains beyond this setting.

翻译：联邦学习通过无需集中聚合数据即可实现协作模型训练，已在关键领域的机器学习应用中占据重要地位。然而，保护隐私的联邦学习框架往往以牺牲公平性和可靠性为代价。差分隐私虽能减少数据泄露，但可能掩盖用于偏差校正的敏感属性，从而加剧不同人口统计组间的性能差距。本文研究了基于联邦学习的目标检测中隐私与公平性的权衡问题，并提出RESFL这一集成框架以协同优化这两个目标。RESFL将对抗性隐私解耦与不确定性引导的公平感知聚合相结合。对抗性组件利用梯度反转层抑制敏感属性信息，在减少隐私风险的同时保留公平相关结构；不确定性感知聚合组件则通过证据神经网络自适应加权客户端更新，优先聚合具有较低公平性差异和较高置信度的贡献，从而生成鲁棒且公平的联邦学习模型更新。在高风险自动驾驶场景下的实验表明：RESFL在FACET和CARLA上实现了高mAP，将成员推断攻击成功率降低了37%，相较于FedAvg基线将机会均等差距缩小了17%，并保持了更强的对抗鲁棒性。尽管在自动驾驶场景中评估，RESFL具有领域无关性，可应用于该场景之外的广泛领域。