BDPFL：基于可解释蒸馏的个性化联邦学习后门防御
BDPFL: Backdoor Defense for Personalized Federated Learning via Explainable Distillation
AI总结：
针对个性化联邦学习易受后门攻击的问题，提出名为BDPFL的防御框架，通过逐层互蒸馏和解释热力图学习高质量中间表示来缓解后门，实验证明其优于基线方法。
摘要：
联邦学习是一种分布式学习范式，它促进了在多个客户端上协作训练全局模型，同时保护本地数据集的隐私。为了解决数据异构性相关的内在挑战并满足个性化需求，联邦学习中的一个新方向，即个性化联邦学习（pFL），逐渐发展起来。人们广泛关注开发新的框架和方法来提升pFL的性能。遗憾的是，pFL的安全性问题在很大程度上被忽视了。我们的目标是填补这一空白。与FL类似，pFL也容易受到后门攻击。然而，现有的后门防御策略主要针对通用FL框架定制，pFL缺乏抵御后门攻击的鲁棒性。我们提出了一种名为BDPFL的新型后门鲁棒pFL框架来应对这些挑战。首先，BDPFL引入了逐层互蒸馏，使客户端能够学习其个性化本地模型，同时缓解潜在的后门。然后，BDPFL采用解释热力图来学习高质量的中间表示，并增强消除更深、更根深蒂固的后门的效果。此外，我们在三个数据集上对BDPFL的性能进行了实证评估，并将BDPFL与四种后门防御方法进行了比较。实验表明，BDPFL优于基线方法，并且在各种设置下均有效。
英文摘要：
Federated learning is a distributed learning paradigm that facilitates the collaborative training of a global model across multiple clients while preserving the privacy of local datasets. To address inherent challenges related to data heterogeneity and satisfy personalized needs, a new direction within FL, known as personalized Federated Learning (pFL), has gradually evolved. Extensive attention has been directed toward developing novel frameworks and methods to enhance the performance of pFL. Regrettably, the aspect of security in pFL has been largely overlooked. Our objective is to fill this gap. Similar to FL, pFL is susceptible to backdoor attacks. However, existing backdoor defense strategies are primarily tailored to general FL frameworks, and pFL lacks robustness against backdoor attacks. We propose a novel, backdoor-robust pFL framework named BDPFL to address these challenges. First, BDPFL introduces layer-wise mutual distillation that enables clients to learn their personalized local models while mitigating potential backdoors. Then, BDPFL employs explanation heatmap to learn high-quality intermediate representations and enhance the effect of eliminating deeper and more entrenched backdoors. Moreover, we perform empirical evaluations of BDPFL's performance on three datasets and compare BDPFL with four backdoor defense methods. The experiments demonstrate that BDPFL outperforms baseline methods and is effective under various settings.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
