Scholarlink

<p>Decentralized Federated Learning enhances privacy by removing central servers, but this architecture is susceptible to model poisoning and sophisticated network-level attacks. The Byzantine-robust BALANCE algorithm is a key defense against malicious updates, but this defense breaks down and halts learning if a network-level attack can feed a node a supermajority of malicious models. To counter this fundamental vulnerability, we introduce RADAR: Robust Aggregation with Dynamic Adversarial-Resilient Routing for such systems. The primary contribution of our RADAR framework is its use of Dynamic Adversarial-Resilient Routing-implemented as a proactive Moving Target Defense to protect its Robust Aggregation core. By creating an unpredictable network topology, RADAR prevents the network-level conditions that disable BALANCE, denying adversaries the sustained dominance required to overwhelm a node. Furthermore, RADAR enhances standard Moving Target Defense by replacing simplistic, deadlock-vulnerable random peer selection with a more secure, adaptive and unpredictable neighbor selection protocol. Communication integrity is further ensured through the use of Transport Layer Security, which provides robust encryption and protection against man-in-the-middle attacks. Importantly, our experimental results confirm that the integration of these techniques simultaneously preserves the efficiency and benefits of the aggregation rule, as demonstrated by an average F1 score of 0.92, representing the most secure configuration under model poisoning attacks-while imposing no significant overhead on system performance, with CPU and RAM usage increasing by only 2-3%, as well as approximately 45% reduced network traffic compared to the baseline.</p>

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

RADAR: Robust aggregation with dynamic adversarial-resilient routing decentralized federated learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月29日

查看全文
  • Duc Dung Truong,
  • Ba Hoang Pham,
  • Dinh Tien Ta,
  • Trong Nhan Phan,
  • Thi Ai Thao Nguyen

摘要

Decentralized Federated Learning enhances privacy by removing central servers, but this architecture is susceptible to model poisoning and sophisticated network-level attacks. The Byzantine-robust BALANCE algorithm is a key defense against malicious updates, but this defense breaks down and halts learning if a network-level attack can feed a node a supermajority of malicious models. To counter this fundamental vulnerability, we introduce RADAR: Robust Aggregation with Dynamic Adversarial-Resilient Routing for such systems. The primary contribution of our RADAR framework is its use of Dynamic Adversarial-Resilient Routing-implemented as a proactive Moving Target Defense to protect its Robust Aggregation core. By creating an unpredictable network topology, RADAR prevents the network-level conditions that disable BALANCE, denying adversaries the sustained dominance required to overwhelm a node. Furthermore, RADAR enhances standard Moving Target Defense by replacing simplistic, deadlock-vulnerable random peer selection with a more secure, adaptive and unpredictable neighbor selection protocol. Communication integrity is further ensured through the use of Transport Layer Security, which provides robust encryption and protection against man-in-the-middle attacks. Importantly, our experimental results confirm that the integration of these techniques simultaneously preserves the efficiency and benefits of the aggregation rule, as demonstrated by an average F1 score of 0.92, representing the most secure configuration under model poisoning attacks-while imposing no significant overhead on system performance, with CPU and RAM usage increasing by only 2-3%, as well as approximately 45% reduced network traffic compared to the baseline.