Scholarlink

Despite the remarkable success and widespread application of deep neural networks (DNNs), the aspect of fairness in DNNs often goes overlooked. In particular, the unfairness threats is more practical and urgent in federated learning (FL), which allows multiple clients to collaboratively train models and greatly facilitates attackers to manipulate models. Surprisingly, this palpable risk has yet to be fully acknowledged and explored in the research community. To bridge this gap, this paper introduces the Federated Unfairness Distribution Attack (FedUDA), a pioneering strategy designed to deliberately induce bias covertly and realistically into FL systems, which poses serious and significant threats in real worlds. FedUDA is achieved by the strategic manipulation of data distribution within a single malicious client, without sacrificing the accuracy of the model’s outputs by leveraging gradient decomposition to isolate and inject bias-inducing updates orthogonal to the main learning objective. It improves durability by focusing on stable neural connections. Our experimental results demonstrate the effectiveness and real threat of the proposed method, and reveal for the first time the vulnerability of FL to unfairness attacks. Our research sheds light on the often-neglected unfairness risks in FL, emphasizing the critical need for the development of stronger safeguards to promote fairness in federated learning.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FedUDA: Towards a Novel Unfairness Distribution Attack Against Federated Learning Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Zihao Wang,
  • Xiaodong Fu,
  • Li Liu,
  • Fei Dai,
  • Jiaman Ding

摘要

Despite the remarkable success and widespread application of deep neural networks (DNNs), the aspect of fairness in DNNs often goes overlooked. In particular, the unfairness threats is more practical and urgent in federated learning (FL), which allows multiple clients to collaboratively train models and greatly facilitates attackers to manipulate models. Surprisingly, this palpable risk has yet to be fully acknowledged and explored in the research community. To bridge this gap, this paper introduces the Federated Unfairness Distribution Attack (FedUDA), a pioneering strategy designed to deliberately induce bias covertly and realistically into FL systems, which poses serious and significant threats in real worlds. FedUDA is achieved by the strategic manipulation of data distribution within a single malicious client, without sacrificing the accuracy of the model’s outputs by leveraging gradient decomposition to isolate and inject bias-inducing updates orthogonal to the main learning objective. It improves durability by focusing on stable neural connections. Our experimental results demonstrate the effectiveness and real threat of the proposed method, and reveal for the first time the vulnerability of FL to unfairness attacks. Our research sheds light on the often-neglected unfairness risks in FL, emphasizing the critical need for the development of stronger safeguards to promote fairness in federated learning.