Privacy and quality of service aware edge network design

Abstract

As the network infrastructure grows increasingly more capable of supporting large throughput, other factors such as privacy and latency remain increasingly important in network design. In this thesis, we target two trade-offs: The Privacy-Rate-Memory trade-off in the edge caching systems and the delivery-latency trade-off in transport layer protocol design. In scenarios where one server uses a multicast channel to communicate with multiple cache-enabled users, coded caching schemes achieve a low server transmission rate by jointly optimizing both the cache content placement of each user cache and the server transmission. We observe that the coded caching scheme achieves a low rate but sacrifices user privacy. An attacker can gain knowledge about user file requests by analyzing server transmissions. This observation is then formally formulated in this thesis as feasible attack algorithms that predict user file requests by either an eavesdropper or colluding users. To address the privacy leakage in coded caching that is noted above, we present two new coded caching schemes that protect user privacy from potential attackers who may be the server or colluding users. Firstly, we propose a scheme that splits the user cache into a private portion to conduct uncoded caching and a public portion to conduct the coded caching scheme. Since there is no privacy issue for uncoded caching, the user protects its privacy by losing some gain in rate. Secondly, we propose another scheme that turns the one-to-one mapping between user file requests and the server transmission into a one-to-many mapping. Therefore, the attacker cannot reversely predict the user file request from the known server transmission. The new mapping is achieved by grouping multiple files into a big file and conducting a coded caching scheme with different file sizes. For both schemes, we study the associated Privacy-Rate-Memory trade-off (PRM). We next consider the design of a quality-of-service aware transport layer for delay and packet loss sensitive applications. We first present the design and performance evaluation of a new Double Q-learning Network based transport layer protocol called RCP-DQN. Unlike prior transport layer protocols that aim at either providing guaranteed end-to-end delivery, e.g., TCP, or minimizing the end-to-end delay, e.g., UDP, RCP-DQN aims to optimize any achievable combination of these objectives, as specified by an application layer utility function. The considered utility function can be quite general and can capture a combination of delay and packet delivery metrics. RCP-DQN can be thought of as an intelligent middle-ground between UDP and TCP that maps application layer objectives subject to what is learned about the network state. It is window-based like TCP, but retransmissions are decided based on a reinforcement learning algorithm to maximize the application layer utility function. RCP-DQN employs a reinforcement learning method, a double Q-Learning network, to learn the best strategy in real-time from a history of packet (re)transmission experiences. No assumption on the shape of the utility function is needed. RCP-DQN is evaluated under a wide range of network settings and is found to outperform UDP, TCP, and ARQ for almost all settings. The performance is also evaluated with respect to TCP-friendliness and network stability. The superior performance of RCP-DQN motivates a systematic analysis of the packet Retransmission Control Problem (RCP). We formulate the RCP problem as a semi-Markov Decision Problem (SMDP) by requiring the utility function to be a time-discounted utility function. Under the SMDP setup, we study the optimal policy (ORCP) and the achieved reward. By making use of the structure of the RCP problem, we show that the optimal policy can be simplified to a control-limit policy (CRCP). We prove that the control-limit policy is still optimal and derive bounds on the performance. For applications, we present a Control-Limit policy with Estimation algorithm (CERCP) by adding network status and system performance estimations. After that, we focus on the relaxation of the requirement on the shape of the utility function. Motivated by the optimal policy, we presented a Q-Learning based policy (QRCP), which has much fewer constraints on the utility function. QRCP follows a customized Q table updating strategy and does not explicitly require the estimation of the network state and system performance. Both policies are evaluated under multiple network settings, and both of them outperform UDP, TCP and ARQ for almost all network settings.