NSF-IITP: CNS Core: Small: Federated Learning for Privacy-preserving Video Caching Network

University Of Southern California, Los Angeles CA

Investigators

Abstract

Video streaming has, over the past decade, become the dominant form of entertainment, with most video clips on social media as well as movies being viewed on wireless devices. Due to large increase in demand, novel methods for content delivery must be considered so as to not overburden the wireless networks, while at the same time keeping transmission delays low, to avoid the dreaded “buffering” warning on the device of the user. One of the most promising methods for achieving this is caching at the wireless edge, i.e., storing popular content at or near the wireless base stations that users are connected to. Caching strategies usually require knowledge of the video popularity, as well as the historical preferences of the individual users. This information needs to be combined with information about the wireless network structure and network parameters, to determine what content should be cached where. Yet, in many cases it is undesirable to share detailed user profiles with the cache and/or wireless network operator. The main motivation for avoiding such sharing is user privacy. The goal of the current project is thus to develop techniques based on machine learning that enable efficient video caching and delivery systems while preserving the privacy of the users. Such algorithms will advance the state of the art in machine learning, by incorporating expert knowledge on video caching and developing new network structures derived from the specific problem. At the same time the results will benefit society by providing privacy – which is of great importance to consumers and steadily becoming more so – as well as spectral and energy efficiency of wireless networks and thus careful use of finite resources. The project will also serve to help students develop interdisciplinary thinking, and has a detailed plan for increasing the participation from underrepresented groups. The project will explore caching techniques that preserve privacy while retaining the efficiency of caching. In particular it will use Federated Learning , a form of Machine Learning that allows localized training and exchange of machine learning models, such as parameterized neural networks, between distributed nodes without requiring the exchange of underlying data such as user preferences. The developed techniques will take into account both content popularity and the state of the wireless network. Rather than separating the problem into disparate parts, such as popularity prediction on one hand, and wireless caching optimization on the other, the project will pursue an integrated approach that accounts for the interaction between all the different aspects of the work. Specific topics of investigation will include (i) techniques that do not require transmission of models to a central server, thus further strengthening privacy, (ii) algorithms for hybrid global-local models that take account of the fact that video popularity is a global descriptor while also showing some local (spatio-temporal) variations, (iii) direct learning of caching strategies without taking the detour via separate learning of video popularity and wireless network state, thus improving both efficiency and privacy. All these results will flow into an integrated, holistic system design for privacy-preserving video caching systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.