Search

An intriguing property of deep neural networks is the existence of adversarial examples, which can transfer among different architectures. These transferable adversarial examples may severely hinder deep neural network-based applications. Previous works mostly study the transferability using small scale datasets. In this work, we are the first to conduct an extensive study of the transferability over large models and a large scale dataset, and we are also the first to study the transferability of targeted adversarial examples with their target labels. We study both non-targeted and targeted adversarial examples, and show that while transferable non-targeted adversarial examples are easy to find, targeted adversarial examples generated using existing approaches almost never transfer with their target labels. Therefore, we propose novel ensemble-based approaches to generating transferable adversarial examples. Using such approaches, we observe a large proportion of targeted adversarial examples that are able to transfer with their target labels for the first time. We also present some geometric studies to help understanding the transferable adversarial examples. Finally, we show that the adversarial examples generated using ensemble-based approaches can successfully attack Clarifai.com, which is a black-box image classification system.

Future servers will incorporate many active lowpower modes for different system components, such as cores and memory. Though these modes provide flexibility for power management via Dynamic Voltage and Frequency Scaling (DVFS), they must be operated in a coordinated manner. Such coordinated control creates a combinatorial space of possible power mode configurations. Given the rapid growth of the number of cores, it is becoming increasingly challenging to quickly select the configuration that maximizes the performance under a given power budget. Prior power capping techniques do not scale well to large numbers of cores, and none of those works has considered memory DVFS. In this paper, we present FastCap, our optimization approach for system-wide power capping, using both CPU and memory DVFS. Based on a queuing model, FastCap formulates power capping as a non-linear optimization problem where we seek to maximize the system performance under a power budget, while promoting fairness across applications. Our FastCap algorithm solves the optimization online and efficiently (low complexity on the number of cores), using a small set of performance counters as input. To evaluate FastCap, we simulate it for a many-core server running different types of workloads. Our results show that FastCap caps power draw accurately, while producing better application performance and fairness than many existing CPU power capping methods (even after they are extended to use of memory DVFS as well)., Comment: Accepted by ISPASS 2016

Power consumption in data centers has been growing significantly in recent years. To reduce power, servers are being equipped with increasingly sophisticated power management mechanisms. Different mechanisms offer dramatically different trade-offs between power savings and performance penalties. Considering the complexity, variety, and temporally varying nature of the applications hosted in a typical data center, intelligently determining which power management policy to use and when is a complicated task. In this paper we analyze a system model featuring both performance scaling and low-power states. We reveal the interplay between performance scaling and low-power states via intensive simulation and analytic verification. Based on the observations, we present SleepScale, a runtime power management tool designed to efficiently exploit existing power control mechanisms. At run time, SleepScale characterizes power consumption and quality-of-service (QoS) for each low-power state and frequency setting, and selects the best policy for a given QoS constraint. We evaluate SleepScale using workload traces from data centers and achieve significant power savings relative to conventional power management strategies., Comment: Accepted by ACM/IEEE International Symposium on Computer Architecture (ISCA) 2014

In this paper we study how coding in distributed storage reduces expected download time, in addition to providing reliability against disk failures. The expected download time is reduced because when a content file is encoded to add redundancy and distributed across multiple disks, reading only a subset of the disks is sufficient to reconstruct the content. For the same total storage used, coding exploits the diversity in storage better than simple replication, and hence gives faster download. We use a novel fork-join queuing framework to model multiple users requesting the content simultaneously, and derive bounds on the expected download time. Our system model and results are a novel generalization of the fork-join system that is studied in queueing theory literature. Our results demonstrate the fundamental trade-off between the expected download time and the amount of storage space. This trade-off can be used for design of the amount of redundancy required to meet the delay constraints on content delivery., Comment: Paper accepted by JSAC-DS 2013

In this paper we study the power-performance relationship of power-efficient computing from a queuing theoretic perspective. We investigate the interplay of several system operations including processing speed, system on/off decisions, and server farm size. We identify that there are oftentimes "sweet spots" in power-efficient operations: there exist optimal combinations of processing speed and system settings that maximize power efficiency. For the single server case, a widely deployed threshold mechanism is studied. We show that there exist optimal processing speed and threshold value pairs that minimize the power consumption. This holds for the threshold mechanism with job batching. For the multi-server case, it is shown that there exist best processing speed and server farm size combinations., Comment: Paper published in CISS 2013

We study the fundamental trade-off between storage and content download time. We show that the download time can be significantly reduced by dividing the content into chunks, encoding it to add redundancy and then distributing it across multiple disks. We determine the download time for two content access models - the fountain and fork-join models that involve simultaneous content access, and individual access from enqueued user requests respectively. For the fountain model we explicitly characterize the download time, while in the fork-join model we derive the upper and lower bounds. Our results show that coding reduces download time, through the diversity of distributing the data across more disks, even for the total storage used., Comment: 8 pages, 6 figures, conference

The vertex v of a graph G is called a 1-critical-vertex for the maximum genus of the graph, or for simplicity called 1-critical-vertex, if G-v is a connected graph and {\deg}M(G - v) = {\deg}M(G) - 1. In this paper, through the joint-tree model, we obtained some types of 1-critical-vertex, and get the upper embeddability of the Spiral Snm .

The weak minor G of a graph G is the graph obtained from G by a sequence of edge-contraction operations on G. A weak-minor-closed family of upper embeddable graphs is a set G of upper embeddable graphs that for each graph G in G, every weak minor of G is also in G. Up to now, there are few results providing the necessary and sufficient conditions for characterizing upper embeddability of graphs. In this paper, we studied the relation between the vertex splitting operation and the upper embeddability of graphs; provided not only a necessary and sufficient condition for characterizing upper embeddability of graphs, but also a way to construct weak-minor-closed family of upper embeddable graphs from the bouquet of circles; extended a result in J: Graph Theory obtained by L. Nebesk{\P}y. In addition, the algorithm complex of determining the upper embeddability of a graph can be reduced much by the results obtained in this paper.

In this paper we develop scheduling policies that maximize the stability region of a wireless network under the assumption that mutual information accumulation is implemented at the physical layer. When the link quality between nodes is not sufficiently high that a packet can be decoded within a single slot, the system can accumulate information across multiple slots, eventually decoding the packet. The result is an expanded stability region. The accumulation process over weak links is temporally coupled and therefore does not satisfy the independent and identically distributed (i.i.d) assumption that underlies many previous analysis in this area. Therefore the problem setting also poses new analytic challenges. We propose two dynamic scheduling algorithms to cope with the non-i.i.d nature of the decoding. The first performs scheduling every $T$ slots, and approaches the boundary of the stability region as $T$ gets large, but at the cost of increased average delay. The second introduces virtual queues for each link and constructs a virtual system wherein two virtual nodes are introduced for each link. The constructed virtual system is shown to have the same stability region as the original system. Through controlling the virtual queues in the constructed system, we avoid the non-i.i.d analysis difficulty and attain the full stability region. We derive performance bounds for both algorithms and compare them through simulation results., Comment: submitted to IEEE Trans. on Information Theory, September 2011; revised, May 2012

We consider the resource allocation problem in cooperative wireless networks wherein nodes perform mutual information accumulation. We consider a unicast setting and arbitrary arrival processes at the source node. Source arrivals can be broken down into numerous packets to better exploit the spatial and temporal diversity of the routes available in the network. We devise a linear-program-based algorithm which allocates network resource to meet a certain transmission objective. Given a network, a source with multiple arriving packets and a destination, our algorithm generates a policy that regulates which nodes should participate in transmitting which packets, when and with what resource. By routing different packets through different nodes the policy exploits spatial route diversity, and by sequencing packet transmissions along the same route it exploits temporal route diversity.

We design and analyze a method to extract secret keys from the randomness inherent to wireless channels. We study a channel model for multipath wireless channel and exploit the channel diversity in generating secret key bits. We compare the key extraction methods based both on entire channel state information (CSI) and on single channel parameter such as the received signal strength indicators (RSSI). Due to the reduction in the degree-of-freedom when going from CSI to RSSI, the rate of key extraction based on CSI is far higher than that based on RSSI. This suggests that exploiting channel diversity and making CSI information available to higher layers would greatly benefit the secret key generation. We propose a key generation system based on low-density parity-check (LDPC) codes and describe the design and performance of two systems: one based on binary LDPC codes and the other (useful at higher signal-to-noise ratios) based on four-ary LDPC codes.

A map is a connected topological graph $\Gamma$ cellularly embedded in a surface. In this paper, applying Tutte's algebraic representation of map, new ideas for enumerating non-equivalent orientable or non-orientable maps of graph are presented. By determining automorphisms of maps of Cayley graph $\Gamma={\rm Cay}(G:S)$ with ${\rm Aut} \Gamma\cong G\times H$ on locally, orientable and non-orientable surfaces, formulae for the number of non-equivalent maps of $\Gamma$ on surfaces (orientable, non-orientable or locally orientable) are obtained . Meanwhile, using reseults on GRR graph for finite groups, we enumerate the non-equivalent maps of GRR graph of symmetric groups, groups generated by 3 involutions and abelian groups on orientable or non-orientable surfaces., Comment: 17 pages with 1 figure

A map is a connected topological graph cellularly embedded in a surface and a complete map is a cellularly embedded complete graph in a surface. In this paper, all automorphisms of complete maps of order n are determined by permutations on its vertices. Applying a scheme for enumerating maps on surfaces with a given underlying graph, the numbers of unrooted complete maps on orientable or non-orientable surfaces are obtained., Comment: 21 pages with 2 figures