The remarkable performance achieved by Deep Neural Networks (DNNs) in many applications is followed by the rising concern about data privacy and security. Since DNNs usually require large datasets to train, many practitioners scrape data from external sources such as the Internet. However, an external data owner may not be willing to let this happen, causing legal or ethical issues. In this paper, we study the generalization attacks against DNNs, where an attacker aims to slightly modify training data in order to spoil the training process such that a trained network lacks generalizability. These attacks can be performed by data owners and protect data from unexpected use. However, there is currently no efficient generalization attack against DNNs due to the complexity of a bilevel optimization involved. We propose the Neural Tangent Generalization Attack (NTGA) that, to the best of our knowledge, is the first work enabling clean-label, black-box generalization attack against DNNs. We conduct extensive experiments, and the empirical results demonstrate the effectiveness of NTGA. Our code and perturbed datasets are available at: https://github.com/lionelmessi6410/ntga.

Object detection based on pre-trained deep neural networks (DNNs) has achieved impressive performance and enabled many applications. However, DNN-based object detectors are shown to be vulnerable to physical adversarial attacks. Despite that recent efforts have been made to defend against these attacks, they either use strong assumptions or become less effective with pre-trained object detectors. In this paper, we propose adversarial pixel masking (APM), a defense against physical attacks, which is designed specifically for pre-trained object detectors. APM does not require any assumptions beyond the "patch-like" nature of a physical attack and can work with different pre-trained object detectors of different architectures and weights, making it a practical solution in many applications. We conduct extensive experiments, and the empirical results show that APM can significantly improve model robustness without significantly degrading clean performance.

Deterministic database systems have been shown to significantly improve the availability and scalability of a distributed database system deployed on a shared-nothing architecture across WAN while ensuring strong consistency. However, their scalability and performance advantages highly depend on the quality of data partitioning due to the reduced flexibility in transaction processing. Although a deterministic database system can employ workload driven data (re-)partitioning and live data migration algorithms to partition data, we found that the effectiveness of these algorithms is limited in complex real-world environments due to the unpredictability of machine workloads. In this paper, we present Hermes, a deterministic database system prototype that, for the first time, does not rely on sophisticated data partitioning to achieve high scalability and performance. Hermes employs a novel transaction routing mechanism that jointly optimizes the balance of machine workloads, data (re-)partitioning, and live data migration by looking into the queued transactions to be executed in the near future. We conducted extensive experiments which show that Hermes is able to yield 29% to 137% increase in transaction throughput as compared to the state-of-the-art systems under complex real-world workloads.

Deep neural networks are shown to be vulnerable to adversarial attacks. This motivates robust learning techniques, such as the adversarial training, whose goal is to learn a network that is robust against adversarial attacks. However, the sample complexity of robust learning can be significantly larger than that of “standard” learning. In this paper, we propose improving the adversarial robustness of a network by leveraging the potentially large test data seen at runtime. We devise a new defense method, called runtime masking and cleansing (RMC), that adapts the network at runtime before making a prediction to dynamically mask network gradients and cleanse the model of the non-robust features inevitably learned during the training process due to the size limit of the training set. We conduct experiments on real-world datasets and the results demonstrate the effectiveness of RMC empirically.

Deep neural networks are shown to be susceptible to both adversarial attacks and backdoor attacks. Although many defenses against an individual type of the above attacks have been proposed, the interactions between the vulnerabilities of a network to both types of attacks have not been carefully investigated yet. In this paper, we conduct experiments to study whether adversarial robustness and backdoor robustness can affect each other and find a trade-off—by increasing the robustness of a network to adversarial examples, the network becomes more vulnerable to backdoor attacks. We then investigate the cause and show how such a trade-off can be exploited for either good or bad purposes. Our findings suggest that future research on defense should take both adversarial and backdoor attacks into account when designing algorithms or robustness measures to avoid pitfalls and a false sense of security.

We study the problem of detecting critical structures using a graph embedding model. Existing graph embedding models lack the ability to precisely detect critical structures that are specific to a task at the global scale. In this paper, we propose a novel graph embedding model, called the Ego-CNNs, that employs the ego-convolutions convolutions at each layer and stacks up layers using an ego-centric way to detects precise critical structures efficiently. An Ego-CNN can be jointly trained with a task model and help explain/discover knowledge for the task. We conduct extensive experiments and the results show that Ego-CNNs (1) can lead to comparable task performance as the state-of-the-art graph embedding models, (2) works nicely with CNN visualization techniques to illustrate the detected structures, and (3) is efficient and can incorporate with scale-free priors, which commonly occurs in social network datasets, to further improve the training efficiency.

The Convolutional Neural Networks (CNNs) have laid the foundation for many techniques in various applications. Despite achieving remarkable performance in some tasks, the 3D viewpoint generalizability of CNNs is still far behind humans visual capabilities. Although recent efforts, such as the Capsule Networks, have been made to address this issue, these new models are either hard to train and/or incompatible with existing CNN-based techniques specialized for different applications. Observing that humans use binocular vision to understand the world, we study in this paper whether the 3D viewpoint generalizability of CNNs can be achieved via a binocular vision. We propose CNN^{2}, a CNN that takes two images as input, which resembles the process of an object being viewed from the left eye and the right eye. CNN^{2} uses novel augmentation, pooling, and convolutional layers to learn a sense of three-dimensionality in a recursive manner. Empirical evaluation shows that CNN^{2} has improved viewpoint generalizability compared to vanilla CNNs. Furthermore, CNN^{2} is easy to implement and train, and is compatible with existing CNN-based specialized techniques for different applications.

Recent deterministic database systems have achieved high scalability and high availability in distributed environments given OLTP workloads. However, modern OLTP applications usually have changing workloads or access patterns, so how to make the resource provisioning elastic to the changing workloads becomes an important design goal for a deterministic database system. Live migration, which moves the specified data from a source machine to a destination node while continuously serving the incoming transactions, is a key technique required for the elasticity. In this paper, we present MgCrab, a live migration technique for a deterministic database system, that leverages the determinism to maintain the consistency of data on the source and destination nodes at very low cost during a migration period. We implement MgCrab on an open-source database system. Extensive experiments were conducted and the results demonstrate the effectiveness of MgCrab.

Cognitive radios (CRs) are proposed to alleviate the huge need for radio spectrum. There are two known steps for a pair of CRs to start communication: the rendezvous and data-channel negotiation. Despite that the rendezvous can be achieved by some well-studied techniques such as the channel hopping, the strategies for data-channel negotiation receive much less attention and their impact on data transmission performance remains unclear. In this paper, we study existing data-channel negotiation schemes for channel-hopping CRs and observe that 1) for short data transmission, they incur a huge overhead, called notification delay, that severely limits the throughput; and 2) for long data transmission, they lead to large overhead, called interruption delay, in handling the PU interruption, which makes performance unstable. By carefully re-examining the steps toward low-overhead and stable data transmission, we argue that a key step, called self-channel selection, is missing and should precede the rendezvous and data channel selection steps. In this step, each CR selects, in a distributed manner, only a small amount of the most stable channels to be used in the later steps. To realize the self-channel selection, we introduce a Randomly-Started Stability-Descent (RSSD) selection algorithm. Expensive simulations are conducted and the results demonstrate the effectiveness of RSSD in reducing 1) the notification delay; 2) the chance of PU interruption; and 3) the interruption delay if PU interruption occurs, which overall improve the performance and quality of data transmission.

Semi-supervised clustering algorithms have been proposed to identify data clusters that align with user perceived ones via the aid of side information such as seeds or pairwise constrains. However, traditional side information is mostly at the instance level and subject to the sampling bias, where non-randomly sampled instances in the supervision can mislead the algorithms to wrong clusters. In this paper, we propose learning from the feature-level supervision. We show that this kind of supervision can be easily obtained in the form of perception vectors in many applications. Then we present novel algorithms, called Perception Embedded (PE) clustering, that exploit the perception vectors as well as traditional side information to find clusters perceived by the user. Extensive experiments are conducted on real datasets and the results demonstrate the effectiveness of PE empirically.

Deterministic database systems have been shown to yield high throughput on a cluster of commodity machines while ensuring the strong consistency between replicas, provided that the data can be well-partitioned on these machines. However, data partitioning can be suboptimal for many reasons in real-world applications. In this paper, we present T-Part, a transaction execution engine that partitions transactions in a deterministic database system to deal with the unforeseeable workloads or workloads whose data are hard to partition. By modeling the dependency between transactions as a T-graph and continuously partitioning that graph, T-Part allows each transaction to know which later transactions on other machines will read its writes so that it can push forward the writes to those later transactions immediately after committing. This forward-pushing reduces the chance that the later transactions stall due to the unavailability of remote data. We implement a prototype for T-Part. Extensive experiments are conducted and the results demonstrate the effectiveness of T-Part.

The Cross Domain Collaborative Filtering (CDCF) exploits the rating matrices from multiple domains to make better recommendations. Existing CDCF methods adopt the sub-structure sharing technique that can only transfer linearly correlated knowledge between domains. In this paper, we propose the notion of Hyper-Structure Transfer (HST) that requires the rating matrices to be explained by the projections of some more complex structure, called the hyper-structure, shared by all domains, and thus allows the non-linearly correlated knowledge between domains to be identified and transferred. Extensive experiments are conducted and the results demonstrate the effectiveness of our HST models empirically.

We study the target node prediction problem: given two social networks, identify those nodes/users from one network (called the source network) who are likely to join another (called the target network, with nodes called target nodes). Although this problem can be solved using existing techniques in the field of cross domain classification, we observe that in many real-world situations the cross-domain classifiers perform sub-optimally due to the heterogeneity between source and target networks that prevents the knowledge from being transferred. In this paper, we propose learning the consistent behavior of common users to help the knowledge transfer. We first present the Consistent Incidence Co-Factorization (CICF) for identifying the consistent users, i.e., common users that behave consistently across networks. Then we introduce the Domain-UnBiased (DUB) classifiers that transfer knowledge only through those consistent users. Extensive experiments are conducted and the results show that our proposal copes with heterogeneity and improves prediction accuracy.

Cognitive radio (CR) is intended to meet the exponentially growing demand for spectrum by allowing for opportunistic utilization of idle legacy channels. Rendezvous, where two radios complete handshaking in an idle channel, is a key step for “stranger” (unknown to each other) CRs to start communication. However, none of existing algorithms guarantee rendezvous for heterogeneous or stranger CRs with different spectrum-sensing capabilities, in spite of the fact that (i) a wide variety of mobile devices are equipped with heterogeneous radios and (ii) there are numerous applications requiring efficient rendezvous for heterogeneous radios/CRs. In this paper, we propose a new channel hopping algorithm, called Heterogeneous Hopping (HH), that guarantees rendezvous without assuming existence of a universal channel set that can be sensed by all radios. HH is realized with a two-layer design that harmonizes the fixed-short-cycle and parity-alignment techniques we propose here, in order to guide CRs to rendezvous in two complementary situations resulting from the different capabilities of mobile wireless devices. To best of our knowledge, HH is the first channel-hopping scheme that guarantees rendezvous between heterogeneous radios. Our in-depth evaluation has shown HH to be significantly faster than simple extensions of existing schemes. Moreover, the latter cannot guarantee successful rendezvous, either.

Quorum-based power-saving (QPS) protocols have been proposed for ad hoc networks (e.g., IEEE 802.11 ad hoc mode) to increase energy efficiency and prolong the operational time of mobile stations. These protocols assign to each station a cycle pattern that specifies when the station should wake up (to transmit/receive data) and sleep (to save battery power). In all existing QPS protocols, the cycle length is either identical for all stations or is restricted to certain numbers (e.g., squares or primes). These restrictions on cycle length severely limit the practical use of QPS protocols as each individual station may want to select a cycle length that is best suited for its own need (in terms of remaining battery power, tolerable packet delay, and drop ratio). In this paper, we propose the notion of hyper quorum system (HQS)-a generalization of QPS that allows for arbitrary cycle lengths. We describe algorithms to generate two different classes of HQS given any set of arbitrary cycle lengths as input. We also describe how to find the optimal cycle length for a station to maximize energy efficiency, subject to certain performance constraints. We then present analytical and simulation results that show the benefits of HQS-based power-saving protocols over the existing QPS protocols. The HQS protocols yield up to 41% improvement in energy efficiency under heavy traffic loads while eliminating more than 90% delay drops under light traffic loads.

Cognitive radio allows radio devices to access the idle spectrum opportunistically, thus alleviates the huge demand for spectrum. Rendezvous, where two radios complete handshaking in an idle channel, is a key step for cognitive radios to start communication. Radios may have the same (homogeneous) or different (heterogeneous) spectrum sensing capabilities. Currently, there is a "gap" between the rendezvous algorithms for homogeneous and heterogeneous cognitive radios---existing homogeneous algorithms incur high delay when applied to heterogeneous radios; while heterogeneous algorithms incur high congestion when applied to homogeneous radios. Since mixtures of these two types of radios appear commonly in practice, it is crucial to bridge the gap between the respective rendezvous algorithms. In this paper, we propose a new rendezvous algorithm, named the ICH scheme, for arbitrary mixtures of radios with homogeneous or heterogeneous spectrum sensing capabilities. Rigorous analysis and extensive simulations are conducted and show that ICH is the first rendezvous scheme that guarantees rendezvous for arbitrary mixtures of homogeneous and heterogeneous radios without incurring large delay and congestion.

The Support Vector Machines (SVMs) have been widely used for classification due to its ability to give low generalization error. In many practical applications of classification, however, the wrong prediction of a certain class is much severer than that of the other classes, making the original SVM unsatisfactory. In this paper, we propose the notion of Asymmetric Support Vector Machine (ASVM), an asymmetric extension of the SVM, for these applications. Different from the existing SVM extensions such as thresholding and parameter tuning, ASVM employs a new objective that models the imbalance between the costs of false predictions from different classes in a novel way such that user tolerance on false-positive rate can be explicitly specified. Such a new objective formulation allows us of obtaining a lower false-positive rate without much degradation of the prediction accuracy or increase in training time. Furthermore, we show that the generalization ability is preserved with the new objective. We also study the effects of the parameters in ASVM objective and address some implementation issues related to the Sequential Minimal Optimization (SMO) to cope with large-scale data. An extensive simulation is conducted and shows that ASVM is able to yield either noticeable improvement in performance or reduction in training time as compared to the previous arts.

Asynchronous wakeup schemes have been proposed for ad hoc networks to increase the energy efficiency of wireless communication. The basic idea is to allow a node to sleep when it is idle, and wakeup periodically to check if there are pending transmissions. In this paper, we examine the applicability of asynchronous wakeup schemes to the Mobile Ad Hoc NETworks (MANETs). We discover that, although it is desirable to have nodes with lower mobility to sleep more in reaction to the less-changing link states, in practice this is prohibited due to an unwanted tradeoff between the energy saving and in-time link discovery. All nodes in a network must stay awake frequently based on their highest possible moving speed to avoid network partition. To address this problem, we propose a new wakeup scheme, named Unilateral- (Uni-) scheme, for MANETs that allows nodes with slower moving speed to sleep more without losing the network connectivity. The Uni-scheme supports both the entity mobility and group mobility of nodes, thus has broad applicability. Theoretical analysis and simulation are conducted and show that the Uni-scheme can render significant energy saving as compared with the previous arts.

The problem of coverage is a fundamental issue in wireless sensor networks. In this article, we consider two subproblems: k -coverage contour evaluation and k -coverage rate deployment. The former aims to evaluate, up to k , the coverage level of any location inside a monitored area, while the latter aims to determine the locations of a given set of sensors to guarantee the maximum increment of k -coverage rate when they are deployed into the area. For the k -coverage contour evaluation problem, a nonuniform-grid-based approach is proposed. We prove that the computation cost of our approach is at most the square root of existing solutions. Based on our k -coverage contour evaluation scheme, a greedy k -coverage rate deployment scheme ( k -CRD) is proposed, which is shown to be an order faster than existing studies for k -coverage rate deployment. The k -CRD can incorporate two different heuristics to further reduce its running time. Simulation results show that k -CRD with these heuristics can be significantly more time efficient without causing much degradation in the coverage rate of final deployment.

Location-based channel access protocols have been proposed as a means to broadcast safety related messages through inter-vehicle communications. The protocols divide the road into fixed-size cells and assign a channel to each cell. To broadcast, a vehicle would use the channel assigned to the cell it is currently traveling within. To improve bandwidth utilization, a vehicle may acquire channels dynamically from adjacent cells that are not occupied by other vehicles. In a TDMA setting where each channel is a time slot, message delay occurs as the vehicle must wait for the arrival of the next time slot it owns. This message delay time depends heavily on the adopted cell-to-channel mapping function. We examine an existed naive channel allocation scheme and proposed three new ones. An analysis shows that the proposed schemes may reduce the delay by 50% to 90% .

Asynchronous wakeup schemes have been proposed for ad hoc networks to increase the energy efficiency of wireless communication. The basic idea is to allow a node to sleep when it is idle, and wakeup periodically to check if there are pending transmissions. In this paper we examine the applicability of asynchronous wakeup schemes to the Mobile Ad Hoc Networks (MANETs). We discover that, although it is desirable to have nodes with lower mobility to sleep more in reaction to the less-changing link states, in practice this is prohibited due to an unwanted tradeoff between the energy saving and in-time link discovery. All nodes in a network must stay awake frequently based on their highest possible moving speed in order to avoid network partition. To address this problem, we propose a new wakeup scheme, named Unilateral- (Uni-) scheme, for MANETs that allows nodes with slower moving speed to sleep more without losing the network connectivity. Theoretical analysis shows that the Uni-scheme can render up to 24\% improvement in energy saving as compared with the previous arts.

Clustering in wireless ad hoc networks has shown to be a promising technique to ensure the scalability and efficiency of various communication protocols. Since stations in these networks are usually equipped with batteries as the power source, it is critical to ensure the energy efficiency of a clustering scheme. The Quorum-based Power Saving (QPS) protocols are widely studied over the past years, as they render extensive energy conservation comparing to the IEEE 802.11 Power Saving (PS) mode. However, most existing QPS protocols adopt a symmetric design where each pair of stations in a network are guaranteed to discover each other. Observing that in clustered environments there is no need to insist on all-pair neighbor discovery, we propose an Asymmetric Cyclic Quorum (ACQ) system. The ACQ system guarantees the neighbor discovery between each member node and the clusterhead in a cluster, and between clusterheads in the network. We show that by taxing slightly more energy consumption on the clusterhead, the average energy consumption of stations in a cluster can reduce substantially than can be achieved by traditional QPS protocols. A novel construction scheme is proposed in this work, which assembles the ACQ system in O(1) time. The constructing scheme is adaptive. Stations in a cluster can adjust their awake/sleep ratio collaboratively to strike the balance between energy efficiency and delay under various cluster conditions. Simulation results show that the ACQ system outperforms the previous studies up to 52% in energy efficiency, while introducing no extra worst-case latency.

Intelligent Transportation Systems (ITS) improve passenger/pedestrian safety and transportation productivity through the use of vehicle-to-vehicle and vehicle-to-roadside wireless communication technologies. Communication protocols in these environments must meet strict delay requirements due to the high moving speed of the vehicles. In this paper, we propose an energy-conservative MAC layer protocol, named DSRC-AA, based on IEEE 802.11 that provides power saving to the ITS communication modules (e.g., On Board Units, portable devices, and Road Side Units) while ensuring the bounded delay. DSRC-AA, a generalization of the Asynchronous Quorum-based Power-Saving (AQPS) protocols, capitalizes on the clustering nature of moving vehicles and assigns different wake-up/sleep schedules to the clusterhead and the members of a cluster. DSRC-AA is able to dynamically adapt the schedules to meet the communication delay requirements at various vehicle moving speed. Simulation results show that DSRC-AA is able to yield more than 44 percent reduction in average energy consumption as compared with the existing AQPS protocols, if to be used in vehicular networks.

The Quorum-based Power Saving (QPS) protocols have been proposed to increase the energy efficiency of wireless communication. However, it remains challenging to apply existing QPS protocols to the Mobile Ad Hoc Networks (MANETs) as the timers of nodes are usually asynchronous, the incurred delay are expected to be adaptive, and the network topology is asymmetric. In this paper, we propose an Asynchronous, Adaptive, and Asymmetric (AAA) power management protocol that fulfills the unique requirements of MANETs. We present the asymmetric grid quorum system, a generalization of traditional grid-based quorum systems, to ensure the network connectivity. Theoretical analysis is conducted to demonstrate the benefits of AAA over previous arts.

Quorum-based power saving (QPS) protocols have been proposed for ad hoc networks (e.g., IEEE 802.11 ad hoc mode) to increase energy efficiency and prolong the operational time of mobile stations. These protocols assign to each station a cycle pattern that specifies when the station should wake up (to transmit/receive data) and sleep (to save battery power). In all existing QPS protocols, the cycle length is either identical for all stations or is restricted to certain numbers (e.g. squares or primes). These restrictions on cycle length severely limit the practical use of QPS protocols as each individual station may want to select a cycle length that is best suited for its own need (in terms of remaining battery power, tolerable packet delay, and drop ratio). In this paper we propose the notion of hyper quorum system (HQS)-a generalization of QPS that allows for arbitrary cycle lengths. We describe algorithms to generate two different classes of HQS given any set of arbitrary cycle lengths as input. We then present analytical and simulation results that show the benefits of HQS-based power saving protocols over the existing QPS protocols.

Many practical applications of classification require the classifier to produce a very low false-positive rate. Although the Support Vector Machine (SVM) has been widely applied to these applications due to its superiority in handling high dimensional data, there are relatively little effort other than setting a threshold or changing the costs of slacks to ensure the low false-positive rate. In this paper, we propose the notion of Asymmetric Support VectorMachine (ASVM) that takes into account the false-positives and the user tolerance in its objective. Such a new objective formulation allows us to raise the confidence in predicting the positives, and therefore obtain a lower chance of false-positives. We study the effects of the parameters in ASVM objective and address some implementation issues related to the Sequential Minimal Optimization (SMO) to cope with large-scale data. An extensive simulation is conducted and shows that ASVM is able to yield either noticeable improvement in performance or reduction in training time as compared to the previous arts.

The K-nearest neighbors (KNN) query has been of significant interest in many studies and has become one of the most important spatial queries in mobile sensor networks. Applications of KNN queries may include vehicle navigation, wildlife social discovery, and squad/platoon searching on the battlefields. Current approaches to KNN search in mobile sensor networks require a certain kind of indexing support. This index could be either a centralized spatial index or an in-network data structure that is distributed over the sensor nodes. Creation and maintenance of these index structures, to reflect the network dynamics due to sensor node mobility, may result in long query response time and low battery efficiency, thus limiting their practical use. In this paper, we propose a maintenance-free itinerary-based approach called density-aware itinerary KNN query processing (DIKNN). The DIKNN divides the search area into multiple cone-shape areas centered at the query point. It then performs a query dissemination and response collection itinerary in each of the cone-shape areas in parallel. The design of the DIKNN scheme takes into account several challenging issues such as the trade-off between degree of parallelism and network interference on query response time, and the dynamic adjustment of the search radius (in terms of number of hops) according to spatial irregularity or mobility of sensor nodes. To optimize the performance of DIKNN, a detailed analytical model is derived that automatically determines the most suitable degree of parallelism under various network conditions. This model is validated by extensive simulations. The simulation results show that DIKNN yields substantially better performance and scalability over previous work, both as kappa increases and as the sensor node mobility increases. It outperforms the second runner with up to a 50 percent saving in energy consumption and up to a 40 percent reduction in query response time, while rendering the same level of query result accuracy.

Current approaches to k nearest neighbor (KNN) search in mobile sensor networks require certain kind of indexing support. This index could be either a centralized spatial index or an in-network data structure that is distributed over the sensor nodes. Creation and maintenance of these index structures, to reflect the network dynamics due to sensor node mobility, may result in long query response time and low battery efficiency, thus limiting their practical use. In this paper, we propose a maintenance-free, itinerary-based approach called density-aware itinerary KNN query processing (DIKNN). The DIKNN divides the search area into multiple cone-shape areas centered at the query point. It then performs a query dissemination and response collection itinerary in each of the cone-shape areas in parallel. The design of the DIKNN scheme also takes into account challenging issues such as the the dynamic adjustment of the search radius (in terms of number of hops) according to spatial irregularity or mobility of sensor nodes. The simulation results show that DIKNN yields substantially better performance and scalability over previous work, both as k increases and as the sensor node mobility increases. It outperforms the second runner with up to 50% saving in energy consumption and up to 40% reduction in query response time, while rendering the same level of query result accuracy.