[论文解读] FGAN: Federated Generative Adversarial Networks for Anomaly Detection in Network Traffic

Sankha Das|arXiv (Cornell University)|Mar 21, 2022

Network Security and Intrusion Detection被引用 5

一句话总结

该论文提出FGAN，一种用于大规模网络流量异常检测的联邦生成对抗网络架构，支持跨分布式节点的隐私保护协作模型训练。通过结合联邦学习与加权模型聚合，并采用同态加密技术，FGAN在保护数据隐私的同时，提升了对各类攻击的检测准确率，并降低了模型篡改风险。

ABSTRACT

Over the last two decades, a lot of work has been done in improving network security, particularly in intrusion detection systems (IDS) and anomaly detection. Machine learning solutions have also been employed in IDSs to detect known and plausible attacks in incoming traffic. Parameters such as packet contents, sender IP and sender port, connection duration, etc. have been previously used to train these machine learning models to learn to differentiate genuine traffic from malicious ones. Generative Adversarial Networks (GANs) have been significantly successful in detecting such anomalies, mostly attributed to the adversarial training of the generator and discriminator in an attempt to bypass each other and in turn increase their own power and accuracy. However, in large networks having a wide variety of traffic at possibly different regions of the network and susceptible to a large number of potential attacks, training these GANs for a particular kind of anomaly may make it oblivious to other anomalies and attacks. In addition, the dataset required to train these models has to be made centrally available and publicly accessible, posing the obvious question of privacy of the communications of the respective participants of the network. The solution proposed in this work aims at tackling the above two issues by using GANs in a federated architecture in networks of such scale and capacity. In such a setting, different users of the network will be able to train and customize a centrally available adversarial model according to their own frequently faced conditions. Simultaneously, the member users of the network will also able to gain from the experiences of the other users in the network.

研究动机与目标

• 为解决集中式机器学习在网络安全入侵检测中的局限性，特别是数据隐私保护与跨多样化攻击类型的模型泛化能力问题。
• 构建一个可扩展的、隐私保护的框架，用于在不集中原始流量数据的前提下，在大规模网络中训练鲁棒的异常检测模型。
• 在保持数据机密性和模型完整性的同时，实现分布式网络节点间的协作模型优化。
• 通过同态加密技术，减轻联邦模型聚合过程中模型投毒与篡改攻击的风险。
• 支持检测模型对不同网络区域中不断演变的攻击模式实现动态适应。

提出的方法

• 该框架采用三级联邦学习架构：训练集群、代理服务器和中央服务器，用于协调模型更新。
• 每个训练集群在其本地网络流量上训练一个GAN，其中生成器学习合成真实感流量，判别器学习区分真实与伪造流量。
• 集群的模型更新被发送至代理服务器，后者基于请求优先级和数据质量指标，应用加权平均后再转发至中央服务器。
• 中央服务器使用加密梯度的加权平均执行全局模型聚合，通过同态加密确保隐私保护。
• 在整个训练过程中，所有模型参数和更新均保持加密状态，防止未授权实体访问原始数据或梯度信息。
• 最终聚合的模型被重新分发至所有集群，使所有节点能够从集体学习中受益，同时保护本地数据隐私。

实验结果

研究问题

• RQ1 如何有效结合联邦学习与GAN，以在保护数据隐私的前提下提升大规模网络流量中的异常检测性能？
• RQ2 当更新来自多样化且可能不可信的网络集群时，应采用何种机制确保模型更新聚合的公平性与完整性？
• RQ3 在网络安全入侵检测场景下，如何减轻联邦模型聚合过程中模型篡改与投毒攻击的风险？
• RQ4 基于优先级的更新加权与加密聚合对模型收敛性与检测准确率有何影响？
• RQ5 与集中式GAN相比，联邦GAN架构是否能在多样化攻击类型上实现更好的泛化能力？

主要发现

• 所提出的FGAN框架可在不共享原始流量数据的前提下，实现分布式网络节点间基于GAN的异常检测器协同训练，有效保护隐私。
• 通过基于请求优先级的加权模型聚合机制，系统显著提升了模型更新集成的公平性与鲁棒性。
• 对模型更新实施同态加密，确保中央服务器在不解密单个梯度的情况下完成模型聚合，大幅降低数据泄露或被篡改的风险。
• 通过代理服务器分担聚合任务，减轻了中央服务器的负载，提升了大规模网络环境下的可扩展性。
• 由于多个网络集群提供了多样化的流量特征，最终聚合模型在不同攻击模式下展现出更优的泛化能力。
• 通过端到端加密且可验证的更新机制，系统维持了模型完整性，有效抵御选择性投毒攻击。