In this paper, we consider hybrid quantum key distribution (QKD) networks with primarily quantum repeaters and a small number of trusted nodes. While the trusted nodes need to be trusted, the use of such nodes, together with efficient routing algorithms, can significantly improve the key generation rate. We show that when trusted nodes are placed at asymmetric locations relative to Alice and Bob, however, existing routing algorithms can lead to low key rate. To address this issue, we develop dynamic routing strategies that adjust routing decisions based on the current key pool conditions in the network. In addition, we investigate new and existing classical post-processing techniques that complement the dynamic routing strategies. Using extensive simulations, we show that our dynamic routing strategies can significantly outperform static strategies, and the post-processing techniques are beneficial in high noise scenarios. In addition, combining the dynamic routing strategies and post-processing techniques can further improve the overall key rate.

While routing in wireless networks has been studied extensively, existing protocols are typically designed for a specific set of network conditions and so do not easily accommodate changes in those conditions. For instance, protocols that assume network connectivity cannot be easily applied to disconnected networks. In this paper, we develop a distributed routing strategy based on deep reinforcement learning that generalizes to diverse traffic patterns, congestion levels, network connectivity, and link dynamics. We make the following key innovations in our design: (i) the use of relational features as inputs to the deep neural network approximating the decision space, which enables our algorithm to generalize to diverse network conditions, (ii) the use of packet-centric decisions to transform the routing problem into an episodic task by viewing packets, rather than wireless devices, as reinforcement learning agents, which provides a natural way to propagate and model rewards accurately during learning, and (iii) the use of extended-time actions to model the time spent by a packet waiting in a queue, which reduces the amount of training data needed and allows the learning algorithm to converge more quickly. We evaluate our routing algorithm using a packet-level simulator and show that the policy our algorithm learns during training is able to generalize to larger and more congested networks, different topologies, and diverse link dynamics. Our algorithm outperforms shortest path and backpressure routing with respect to packets delivered and delay per packet.

Consider a multi-hop wireless network in which devices act as anonymizing routers. Even if devices anonymize their link transmissions, an adversary may still be able to infer key information by observing the traffic patterns in the network. In this work, we quantify what impacts how well a Kalman-filter based adversary can infer unlinkability, that is, the probability that different pairs of devices are communicating, from anonymized link transmissions. We assume that devices do not reorder packets to mix traffic and thereby increase unlinkability. Instead, we show that traffic mixing is still possible due to the use of multi-hop routing and broadcast transmissions, with the amount of mixing dependent on the network characteristics. In simulation, we find that i) for unicast links, as network connectivity increases, unlinkability decreases, while for broadcast links as connectivity increases unlinkability increases, ii) link dynamics increase unlinkability in poorly connected topologies, iii) well-connected topologies achieve the same level of unlinkability with fewer transmissions per packet delivered, and (iv) a lattice topology has consistently good unlinkability in different scenarios.

Decoy routing is a powerful circumvention mechanism intended to provide secure communications that cannot be monitored, detected, or disrupted by a third party who controls the user's network infrastructure. Current decoy routing protocols have weaknesses, however: they either make the unrealistic assumption that routes through the network are symmetric (i.e., the router implementing the decoy routing protocol must see all of the traffic, in both directions, from each connection it modifies), or their protocol requires modifying the route taken by packets in connections that use the protocol, and these route changes are detectable by a third party. We present Rebound, a decoy routing protocol that tolerates asymmetric routes without modifying the route taken by any packet that passes through the decoy router, making it more difficult to detect or disrupt than previous decoy routing protocols.

Structural change and uncertainty are fundamental properties of an ad hoc network, making it difficult to develop communication strategies, i.e., network-level approaches to transport data from sender to receiver. At a basic level, change and uncertainty affect how long any state maintained by a communication strategy remains useful, and so influence the trade-offs made to collect that state. In this paper, we introduce a framework for organizing the decision space for deciding when a communication strategy should maintain state, and what type of state should be maintained, in an ad hoc network. The framework is based on our observation that three network properties (connectivity, unpredictability, and resource contention) determine when state is useful. Using the framework, we make three contributions. First, we illustrate the framework by showing an instantiation in terms of specific measures that can be used to describe a network setting. Second, we validate the framework by showing it correctly and consistently organizes the decision space for different communication strategies. Finally, we demonstrate the analytic power of the framework by using it to (1) uncover surprising aspects of well-known traces, and (2) identify the need for, and value of, a new strategy for network communication.

Sensor networks are prone to congestion due to bursty and high-bandwidth data traffic, combined with wireless links and many-to-one data routing to a sink. Delayed and dropped packets then degrade the performance of the sensing application. In this paper, we investigate the value of separate handling of sensor control and data traffic, during times of congestion, in a closed-loop sensor network. We first show that prioritizing sensor control traffic over data traffic decreases the round-trip control-loop delay, and consequently increases the quantity and quality of the data collected by the sensor network. We then ground our analysis in a closed-loop meteorological sensor network, focusing on a storm-tracking application running over a network of X-band radars. Our application measures reflectivity (a measure of the number of scatterers in a unit volume of atmosphere known as a voxel) and tracks storms (i.e., regions of high reflectivity) using a Kalman filter. Considering data quantity, we show that prioritizing sensor control traffic increases the number of voxels, V, that can be scanned given a constant number of reflectivity samples, Nc, obtained per voxel. Here, utility increases linearly with the number of scanned voxels. Considering data quality, we show that prioritizing sensor control traffic increases the number of reflectivity samples, N, that can be obtained per voxel given a constant number of voxels, Vc, to scan. Here, since sensing accuracy improves only as a function of radicN, the gain in accuracy for the reflectivity estimate per voxel as N increases is relatively small except when prioritizing sensor control increases N significantly (such as when sensor control packets suffer severe delays). Because accuracy also degrades as a function of radicN, however, and because prioritizing sensor control traffic reduces the number of control packets dropped, data degradation is mitigated. Considering the performance of the tracking application, we then show that during times of severe congestion, not prioritizing sensor control can actually lead to tracking errors accumulating over time.

In this work we consider the problem of routing in networks with time-varying links. We focus on wireless and mobile ad-hoc networks where link changes occur at the timescale of seconds to minutes. Since links are changing, to maintain network connectivity paths must be periodically recomputed, incurring increased control overhead. Suppose, however, we can perform routing in such a way that paths are "robust" to link changes in the network: i.e., link changes may occur, but paths still perform well enough. Then paths can be recomputed less frequently and the control overhead decreased. In this work we specifically examine the problem of how to perform such "robust" routing.

We address the problem of adaptive sensor control in dynamic resource-constrained sensor networks. We focus on a meteorological sensing network comprising radars that can perform sector scanning rather than always scanning 360 degrees. We compare three sector scanning strategies. The sit-and-spin strategy always scans 360 degrees. The limited lookahead strategy additionally uses the expected environmental state K decision epochs in the future, as predicted from Kalman filters, in its decision-making. The full lookahead strategy uses all expected future states by casting the problem as a Markov decision process and using reinforcement learning to estimate the optimal scan strategy. We show that the main benefits of using a lookahead strategy are when there are multiple meteorological phenomena in the environment, and when the maximum radius of any phenomenon is sufficiently smaller than the radius of the radars. We also show that there is a trade-off between the average quality with which a phenomenon is scanned and the number of decision epochs before which a phenomenon is rescanned.

197-206

This paper investigates learning hierarchical statistical activity models in indoor environments. The abstract hidden Markov model (AHMM) is used to represent behaviors in stochastic environments. We train the model using both labeled and unlabeled data and estimate the parameters using expectation maximization (EM). Results are shown on three datasets: data collected in lab, entryway, and home environments. The results show that hierarchical models outperform flat models.