Year: 2024
Type: dissertation
Abstract: Applications of machine learning to networking, from performance diagnosis to security, have conventionally relied on models that are trained on offline packet traces, without regard to neither the cost of gathering, computing, and storing the corresponding input features nor the performance of the model in its final deployment environment. As a result, there remains a significant gap between the development of statistical models for network operations and their application and systemization in practice. From data collection to feature engineering, from model training to deployment, we argue that all steps of the machine learning development pipeline contain challenges that can hinder the effectiveness of the final model, beyond its accuracy at testing time. The goal of our research is to build new foundational building blocks necessary to address the challenges that emerge when applying machine learning on network traffic deployed in operational networks. In this manuscript, we present our initial work towards the development of new techniques that make it easier and more effective to develop models that work in real-world network deployments. As a first step, we develop new models to infer quality metrics (i.e., startup delay and resolution) for encrypted streaming video services. We demonstrate the models are practical through a 16-month deployment in 66 homes and provide new insights about the relationships between Internet "speed" and the quality of the corresponding video streams, for a variety of services. Building on the lessons learned from this experience, we build solutions that address various challenges across the model development pipeline: (1) To address the lack of labeled data to train models on, we develop a new technique to generate synthetic traffic for training models. (2) To address the need of designing models that account for both predictive performance and systems costs, we develop new techniques to explore cost-aware data representations and automatically train models. (3) To address the challenge of maintaining the accuracy of models in the face of changing data distributions, we develop a new methodology to mitigate model performance decay over time. Overall, our work demonstrates the benefits of systematically addressing the challenges that arise when applying machine learning to network traffic, reducing the gap that limits network operators from deploying models into production.
Source: HAL
Author: Bronzino, Francesco
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4403515608
