The limited availability of high-quality computer networking data, and the privacy risks of sharing what does exist, has prompted development of ML-based methods for generating synthetic network data that mimics real communication between networked devices. The viability of these models hinges on both the quality of their output and how well they preserve private information encoded in their training data. Prior work has sought to address this by training models with differential privacy (DP). However, how this choice affects the actual privacy of the training data, and subsequently the quality of the generated output, is not well understood. In this work, we analyze the relationship between privacy and quality in generative network data models. Using the success of membership inference attacks (MIAs) as the metric for privacy, we observe that whether DP mitigates MIAs depends heavily on model architecture and representation of network data used for training. In particular, we empirically find that some approaches to generating synthetic network data train models that heavily skew towards either overgeneralizing or undergeneralizing to their training data, resulting in poor or inconsistent MIA performance. In these cases, using DP does not yield substantive improvements in vulnerability to MIAs. As for the quality of generated data, we find that DP synthetic network data can retain statistical similarity to real data even under strict privacy budgets, and that downstream models (e.g., classifiers, regressors) trained on this data tend to achieve at least as good accuracy as models trained on non-DP data. These results suggest that DP, depending on the model, offers protection against MIAs without degrading the utility of the generated output, and in some cases, improves utility.

Video conferencing applications (VCAs) have become a critical Internet application during the COVID-19 pandemic, as users worldwide now rely on them for work, school, and telehealth. It is thus increasingly important to understand the resource requirements of different VCAs and how they perform under different network conditions, including: how do application-layer performance metrics (e.g., resolution or frames per second) vary under different link capacity; how VCAs perform under temporary reductions in available capacity; how they compete with themselves, with each other, and with other applications; and how usage modality (e.g., gallery vs. speaker mode) affects utilization. We study three modern VCAs: Zoom, Google Meet, and Microsoft Teams. Answers to these questions differ substantially depending on VCA. First, the average utilization on an unconstrained link varies between 0.8 Mbps and 1.9 Mbps. Given temporary reduction of capacity, some VCAs can take as long as 50 seconds to recover to steady state. Differences in proprietary congestion control algorithms also result in unfair bandwidth allocations: in constrained bandwidth settings, one Zoom video conference can consume more than 75% of the available bandwidth when competing with another VCA (e.g., Meet, Teams). For some VCAs, client utilization can decrease as the number of participants increases, due to the reduced video resolution of each participant's video stream given a larger number of participants. Finally, one participant's viewing mode (e.g., pinning a speaker) can affect the upstream utilization of other participants.