Publications

Modern networks increasingly rely on machine learning models for real-time insights, including traffic classification, application quality of experience inference, and intrusion detection. However, existing approaches prioritize prediction accuracy without considering deployment constraints or the dynamism of network traffic, leading to potentially suboptimal performance. Because of this, deploying ML models in real-world networks with tight performance constraints remains an open challenge. In contrast with existing work that aims to select an optimal candidate model for each task based on offline information, we propose an online, system-driven approach to dynamically select the best ML model for network traffic analysis. To this end, we present Cruise Control, a system that pre-trains several models for a given task with different accuracy-cost tradeoffs and selects the most appropriate model based on lightweight signals representing the system’s current traffic processing ability. Experimental results using two real-world traffic analysis tasks demonstrate Cruise Control’s effectiveness in adapting to changing network conditions. Our evaluation shows that Cruise Control improves median accuracy by 2.78% while reducing packet loss by a factor of four compared to offline-selected models.

Machine Learning (ML)-based traffic analysis relies on a data processing pipeline consisting of multiple steps that filter, process, and collect statistics, or features from raw network traffic. These steps are typically performed by in-network measurement systems deployed in existing network fabric (e.g., programmable switches) or using off-the-shelf hardware (e.g., commodity servers). In both deployment scenarios, these systems come with limited processing budgets that must be finely tuned to precisely collect the required features. Unfortunately, the ever growing traffic volume on modern networks can exhaust these budgets, ultimately resulting in packet loss. In this paper, we investigate the impact of packet loss on the performance of ML-based traffic analysis systems. As losses introduce bias in the final features set provided to the machine learning model, we hypothesize that they will negatively impact model performance. We evaluate this hypothesis by analyzing the performance of two different ML models—service classification and QoE analysis—trained on a dataset of video flows, and we measure the impact of two different packet loss models: probabilistic and bursty losses. Our results show that sporadic packet loss has little impact on performance. Conversely, bursty losses, which are more common for packet processing systems, can lead to a significant negative impact.

Machine learning techniques are a common solution used to solve a variety of network management tasks. Often, a network administrator chooses the model to deploy based on offline information, such as model performance and system load. Yet, network traffic is inherently dynamic making it hard to select an optimal model that can work throughout ever changing conditions. In this paper, we make the case that, instead of having to select the optimal candidate model based on offline information, systems should adapt based on the network conditions observed. We present a system design that takes as input a set of candidate models and their features, and adaptively select the better configuration as a function of the network and the system conditions.

MAC addresses can be collected by passive observers to track Wi-Fi users. While address randomization neutralizes this threat for devices not yet associated, the problem remains for devices being associated with a WLAN. In this paper, we introduce RoMA, which is an anti-tracking scheme making use of concurrent virtual network interfaces (VIFs). We provide a proof-of-concept implementation of RoMA and show that it has a limited impact on the performance of the devices.