Network tomography plays a crucial role in assessing the operational status of internal links within networks through end-to-end path-level measurements, independently of cooperation from the network infrastructure. However, the accuracy of performance inference in internal network links heavily relies on comprehensive end-to-end path performance data. Most network tomography algorithms employ conventional threshold-based methods to identify congestion along paths, while these methods encounter limitations stemming from network complexities, resulting in inaccuracies such as misidentifying abnormal links and overlooking congestion attacks, thereby impeding algorithm performance. This paper introduces the concept of Additive Congestion Status to address these challenges effectively. Using a framework that combines Adversarial Autoencoders (AAE) with Long Short-Term Memory (LSTM) networks, this approach robustly categorizes (as uncongested, single-congested, or multiple-congested) and quantifies (regarding the number of congested links) the Additive Congestion Status. Leveraging prior path information and capturing spatio-temporal characteristics of probing flows, this method significantly enhances the localization of congested links and the inference of link performance compared to conventional network tomography algorithms, as demonstrated through experimental evaluations.

翻译：网络层析在独立于网络基础设施协作的情况下，通过端到端路径级测量评估网络内部链路的运行状态方面发挥着关键作用。然而，内部网络链路性能推断的准确性在很大程度上依赖于全面的端到端路径性能数据。大多数网络层析算法采用传统的基于阈值的方法来识别路径上的拥塞，但这些方法因网络复杂性而存在局限，导致诸如误判异常链路和忽略拥塞攻击等不准确问题，从而阻碍算法性能。本文引入加性拥塞状态的概念以有效应对这些挑战。该方法采用结合对抗自编码器与长短期记忆网络的框架，能够稳健地对加性拥塞状态进行分类（分为无拥塞、单链路拥塞或多链路拥塞）并量化（关于拥塞链路数量）。通过利用先验路径信息并捕获探测流的时空特征，实验评估表明，相较于传统网络层析算法，该方法显著提升了拥塞链路的定位能力与链路性能推断精度。