We study the Patrol Security Game (PSG), a robotic patrolling problem formulated as an extensive-form Stackelberg game, in which the attacker strategically selects the timing, location, and duration of an attack. The defender’s goal is to compute an infinite-horizon patrolling policy that minimizes the attacker’s expected payoff. By restricting the defender’s strategy to a time-homogeneous first-order Markov chain, we show that PSG can be reformulated as a combinatorial minimax problem. We prove that the optimal strategy under zero-penalty scenarios corresponds to minimizing either the expected hitting time or return time, depending on the attacker’s visibility model. These optimal policies are closed-form and can be computed efficiently. On the other hand, in high-penalty cases, we observe that the patrolling schedule with high randomness can minimize the attacker’s expected gain. However, in general, the minimax objective becomes non-convex. To address this, we introduce a bi-criteria optimization framework that jointly considers the expected maximum reward (EMR) and entropy rate of the patrolling policy. We propose three graph-based algorithms and a deep reinforcement learning model to efficiently balance these two objectives. Each algorithm demonstrates distinct strengths under different configurations, such as varying penalty scales and cost function settings. The extensive experiments on both synthetic and real-world crime datasets validate the effectiveness of our approaches, demonstrating superior performance and scalability compared to state-of-the-art baselines.

Industrial recommender systems are frequently tasked with approximating probabilities for multiple, often closely related, user actions. For example, predicting if a user will click on an advertisement and if they will then purchase the advertised product. The conceptual similarity between these tasks has promoted the use of multi-task learning: a class of algorithms that aim to bring positive inductive transfer from related tasks. Here, we empirically evaluate multi-task learning approaches with neural networks for an online advertising task. Specifically, we consider approximating the probability of post-click conversion events (installs) (CVR) for mobile app advertising on a large-scale advertising platform, using the related click events (CTR) as an auxiliary task. We use an ablation approach to systematically study recent approaches that incorporate both multitask learning and “entire space modeling” which train the CVR on all logged examples rather than learning a conditional likelihood of conversion given clicked. Based on these results we show that several different approaches result in similar levels of positive transfer from the data-abundant CTR task to the CVR task and offer some insight into how the multi-task design choices address the two primary problems affecting the CVR task: data sparsity and data bias. Our findings add to the growing body of evidence suggesting that standard multi-task learning is a sensible approach to modelling related events in real-world large-scale applications and suggest the specific multitask approach can be guided by ease of implementation in an existing system.

We present a fully closed-loop design for an artificial pancreas (AP) that regulates the delivery of insulin for the control of Type I diabetes. Our AP controller operates in a fully automated fashion, without requiring any manual interaction with the patient (e.g., in the form of meal announcements). A major obstacle to achieving closed-loop insulin control are the "unknown disturbances" related to various aspects of a patient's daily behavior, especially meals and physical activity. Such disturbances can significantly affect the patient's blood glucose levels. To handle such uncertainties, we present a data-driven, robust, model-predictive control framework in which we capture a wide range of individual meal and exercise patterns using uncertainty sets learned from historical data. These uncertainty sets are then used in the insulin controller to achieve automated, precise, and personalized insulin therapy. We provide an extensive in silico evaluation of our robust AP design, demonstrating the potential of the approach. In particular, without the benefit of explicit meal announcements, our approach can regulate glucose levels for large clusters of meal profiles learned from population-wide survey data and cohorts of virtual patients, even in the presence of high carbohydrate disturbances.

In this paper we propose a new membership attack method called co-membership attacks against deep generative models including Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs). Specifically, membership attack aims to check whether a given instance x was used in the training data or not. A co-membership attack checks whether the given bundle of n instances were in the training, with the prior knowledge that the bundle was either entirely used in the training or none at all. Successful membership attacks can compromise the privacy of training data when the generative model is published. Our main idea is to cast membership inference of target data x as the optimization of another neural network (called the attacker network) to search for the latent encoding to reproduce x. The final reconstruction error is used directly to conclude whether x was in the training data or not. We conduct extensive experiments on a variety of datasets and generative models showing that: our attacker network outperforms prior membership attacks; co-membership attacks can be substantially more powerful than single attacks; and VAEs are more susceptible to membership attacks compared to GANs.

On-street parking is an essential component of parking infrastructure for smart cities, which allows users to park near their destinations for short term. However, due to limited capacity, saturated on-street parking becomes a serious and widespread problem for urban transportation systems. Greedily searching for an on-street parking spot in a saturated area is often a frustrating task for drivers, and cruising for vacant parking spots results in additional delays and impaired local circulation. With the recent development of networked smart parking meter, real-time city-wide on- street parking information becomes available for more efficient parking management. In this paper, we design an online parking guidance system that recommends parking spots in real-time based on the parking availability prediction. With a receding horizon optimization framework, our solution minimizes the user's driving and walking cost by adapting the spatiotemporally dynamic supply and demand in the local area, significantly reducing parking competitions in a timely manner. We implement and evaluate our solution with a dataset of 13,503,655 parking records collected from 5228 in-ground sensors distributed in the Australian city Melbourne. The evaluation results show that our approach achieves up to 63.8% delay reduction compared with existing solutions.

In this paper we study the following problem: given a set of m sensors that collectively cover a set of n target points with heterogeneous coverage requirements (target j needs to be covered every fjslots), how to schedule the sensor duty cycles such that all coverage requirements are satisfied and the maximum number of sensors turned on at any time slot is minimized. The problem models varied real-world applications in which sensing tasks exhibit high discrepancy in coverage requirements - critical locations often need to be covered much more frequently. We provide multiple algorithms with best approximation ratio of O (log n + log m) for the maximum number of sensors to turn on, and bi-criteria algorithm with (α, β)-approximation factors with high probability, where the number of sensors turned on is an α = O(δ(log (n) + log(m))/β)-approximation of the optimal (satisfying all requirements) and the coverage requirement is a β-approximation; δ is the approximation ratio achievable in an appropriate instance of set multi-cover. When the sensor coverage exhibits extra geometric properties, the approximation ratios can be further improved. We also evaluated our algorithms via simulations and experiments on a camera testbed. The performance improvement (energy saving) is substantial compared to turning on all sensors all the time, or a random scheduling baseline.

Understanding occupancy patterns in a building is very useful to control HVAC systems for improving energy efficiency of the building and occupant comfort. There has been a very little attempt to understand long term occupancy patterns in a commercial space. In this work, we leverage depth sensors (Kinect for XBOX One) to collect occupancy count from an 11,000 square foot commercial space (Bosch office) for nine months. We analyze the collected data and describe key findings that provide deep insights about how a commercial space is used and serve as a guideline to formulate novel and efficient control strategies.

As sensor networks are increasingly deployed for critical applications, reliability and latency guarantee become more important than ever to meet industrial requirements. In this paper, we investigated the impact of link burstiness on stream scheduling using a data trace of 3,600,000 packets collected from an indoor testbed. We demonstrate that a good tradeoff between reliability and latency can be achieved by allocating certain time slots on each link for stream transmissions based on its burst length and frequency distributions. With this observation, we design transmission scheduling and routing algorithms for data streams to meet a specified reliability requirement while minimizing end-to- end latency. For the multi-stream scheduling problem, we prove its NP-hardness and design an algorithm that achieves the reliability guarantee and an O(log n) approximation of minimizing the maximum end-to-end latency for any stream. Trace- driven simulations show that our solution meets specified end-to-end reliability requirements with latency up to 9.18 times less than existing solutions.

Using optical sensors for indoor monitoring has been widely adopted in many smart building applications. An important design problem in this space is to explore the tradeoff between energy consumption and coverage quality. While it is important that the sensors achieve full coverage (i.e., every interesting target point can be monitored by at least one sensors), it is often a waste of energy to keep sensors on all the time as events are typically stochastic and rare and most of the time the sensors are on idle monitoring. In this paper we design efficient sensor duty cycles to ensure that any target point of interest is still covered sufficiently frequently while only a subset of sensors are kept on at any time slot. We denote by the maximum dark length for each target point p as the maximum duration in which p is covered at least once. We formulate two optimization problems: the min max dark length scheduling and the min average dark length scheduling. For both versions we provide efficient, practical algorithms with provable approximation guarantee. The two algorithms have been tested on two real testbed scenarios to evaluate its efficiency and coverage quality.

In this paper we investigate sensor deployment and coverage algorithms for using infrared signals in indoor applications. Infrared signals are directional and reliable signals that have little interference with other electromagnetic signals that are commonly found in the deployment domain such as visible light and wireless radio waves. Since the angle of arrival is used, and line of sight is the main constraint for IR signals, we investigate the problem called robust guarding, i.e., placing emitters to ensure that all points of the domain are robustly covered by two emitters that are from sufficiently different directions. We prove combinatorial upper and lower bounds for the number of emitters needed and prove that finding the minimum number of guards is NP-hard. We show that n/2 guards are always sufficient and sometimes necessary for rectilinear polygons and we provide practical algorithms in general. We also developed a testbed with low cost off-the-shelf infrared (IR) emitters and sensors for indoor device-free localization. We tested the algorithms for using infrared sensors for indoor localization and our system achieves an average accuracy of 11.7 cm in a typical office setting.

This article demonstrates how to use multiple channels to improve communication performance in Wireless Sensor Networks (WSNs). We first investigate multichannel realities in WSNs through intensive empirical experiments with Micaz motes. Our study shows that current multichannel protocols are not suitable for WSNs because of the small number of available channels and unavoidable time errors found in real networks. With these observations, we propose a novel tree-based, multichannel scheme for data collection applications, which allocates channels to disjoint trees and exploits parallel transmissions among trees. In order to minimize interference within trees, we define a new channel assignment problem that is proven NP-complete. Then, we propose a greedy channel allocation algorithm that outperforms other schemes in dense networks with a small number of channels. We implement our protocol, called the Tree-based, Multichannel Protocol (TMCP), in a real testbed. To adjust to networks with link quality heterogeneity, an extension of TMCP is also proposed. Through both simulation and real experiments, we show that TMCP can significantly improve network throughput and reduce packet losses. More important, evaluation results show that TMCP better accommodates multichannel realities found in WSNs than other multichannel protocols.