Biblio

We present DeepPicar, a low-cost deep neural network based autonomous car platform. DeepPicar is a small scale replication of a real self-driving car called DAVE-2 by NVIDIA. DAVE-2 uses a deep convolutional neural network (CNN), which takes images from a front-facing camera as input and produces car steering angles as output. DeepPicar uses the same network architecture—9 layers, 27 million connections and 250K parameters—and can drive itself in real-time using a web camera and a Raspberry Pi 3 quad-core platform. Using DeepPicar, we analyze the Pi 3’s computing capabilities to support end-to-end deep learning based real-time control of autonomous vehicles. We also systematically compare other contemporary embedded computing platforms using the DeepPicar’s CNN-based real-time control workload. We find that all tested platforms, including the Pi 3, are capable of supporting the CNN-based real-time control, from 20 Hz up to 100 Hz, depending on hardware platform. However, we find that shared resource contention remains an important issue that must be considered in applying CNN models on shared memory based embedded computing platforms; we observe up to 11.6X execution time increase in the CNN based control loop due to shared resource contention. To protect the CNN workload, we also evaluate state-of-the-art cache partitioning and memory bandwidth throttling techniques on the Pi 3. We find that cache partitioning is ineffective, while memory bandwidth throttling is an effective solution.

Attacks infiltrating the integrity of vehicular control systems and medical devices have brought to sharp focus the urgency of securing cyber-physical systems. There is a broader discussion about the role of principled
security -aware design and analysis in the development of both modern engineering systems such as the Smartgrid as well as in future systems that use advanced AI and machine learning in safety critical settings.
Although there has been a growing interest in these security in the CPSWeek commun ity (increasing number of security related papers in ICCPS, HSCC, RTAS, HyCons), this body of research remains largely disconnected from the mainstream systems security research (USENIX, Oakland, CCS, NDSS). The CPS community has developed analysis and synthesis algorithms, verification tools, notions of observability and controllability, and have been in the forefront of research on emerging applications. The connections between this body of work and systems security research remain unexplored. 

The goal of this workshop is to advance the science of security in cyberphysical systems by helping bridge this. We plan to bring together the leaders from these two communities in a full day workshop of invited
sessions and panel discussions. Instead of unstructured technical presentations, the speakers and participants will put their research in the context of some broad topics that will help us bridge this gap. Topics of interest include:

• Identify hard open problems for academic research in CPS security
• Data and testbeds in security research amenable to CPS methods
• Success and fails in designing for resiliency
• Identify CPS tools and techniques (e.g., verification, synthesis) that can advance systems security research
• How to make an impact with CPS security research (where most systems are closed, design cycles are long, and methodologies are slower to change than in cyber systems)
• Metrics for CPS security

In this work we explore how different cognitive processes af- fected typing patterns through a computer game we call The Typing Game. By manipulating the players’ familiarity with the words in our game through their similarity to dictionary words, and by allowing some players to replay rounds, we found that typing speed improves with familiarity with words, and also with practice, but that these are independent of the number of mistakes that are made when typing. We also found that users who had the opportunity to replay rounds exhibited different typing patterns even before replaying the rounds. 

Bot detection - identifying a software program that's using a computer system -- is an increasingly necessary security task. Existing solutions balance proof of human identity with unobtrusiveness in users' workflows. Cognitive modeling and natural interaction might provide stronger security and less intrusiveness.

The emerging software-defined networking (SDN) technology decouples the control plane from the data plane in a computer network with open and standardized interfaces, and hence opens up the network designers’ options and ability to innovate. The wide adoption of SDN in industry has motivated the development of large-scale, high-fidelity testbeds for evaluation of systems that incorporate SDN. In this article, we develop a framework to support OpenFlow-based SDN simulation and distributed emulation, by leveraging our prior work on a hybrid network testbed with a parallel network simulator and a virtual-machine-based emulation system. We show how to exploit typical SDN controller behaviors to handle performance issues caused by the centralized controller in parallel discrete-event simulation. In particular, we develop an asynchronous synchronization algorithm for passive SDN controllers and design a two-level architecture for active SDN controllers. We evaluate the system performance, showing good scalability. Finally, we present a case study, using the testbed, to evaluate network verification applications in an SDN-based data center network. CCS Concepts: Networks→Network simulations; Computing methodologies→Simulation

As mobile technology begins to dominate computing, understanding how their use impacts security becomes increasingly important. Fortunately, this challenge is also an opportunity: the rich set of sensors with which most mobile devices are equipped provide a rich contextual dataset, one that should enable mobile user behavior to be modeled well enough to predict when users are likely to act insecurely, and provide cognitively grounded explanations of those behaviors. We will evaluate this hypothesis with a series of experiments designed first to confirm that mobile sensor data can reliably predict user stress, and that users experiencing such stress are more likely to act insecurely.

Detecting and preventing attacks before they compromise a system can be done using acceptance testing, redundancy based mechanisms, and using external consistency checking such external monitoring and watchdog processes. Diversity-based adjudication, is a step towards an oracle that uses knowable behavior of a healthy system. That approach, under best circumstances, is able to detect even zero-day attacks. In this approach we use functionally equivalent but in some way diverse components and we compare their output vectors and reactions for a given input vector. This paper discusses practical relevance of this approach in the context of recent web-service attacks.

Techniques commonly used for analyzing streaming video, audio, SIGINT, and network transmissions, at less-than-streaming rates, such as data decimation and ad-hoc sampling, can miss underlying structure, trends and specific events held in the data[3]. This work presents a secure-by-construction approach [7] for the upper-end data streams with rates from 10- to 100 Gigabits per second. The secure-by-construction approach strives to produce system security through the composition of individually secure hardware and software components. The proposed network processor can be used not only at data centers but also within networks and onboard embedded systems at the network periphery for a wide range of tasks, including preprocessing and data cleansing, signal encoding and compression, complex event processing, flow analysis, and other tasks related to collecting and analyzing streaming data. Our design employs a four-layer scalable hardware/software stack that can lead to inherently secure, easily constructed specialized high-speed stream processing. This work addresses the following contemporary problems: (1) There is a lack of hardware/software systems providing stream processing and data stream analysis operating at the target data rates; for high-rate streams the implementation options are limited: all-software solutions can't attain the target rates[1]. GPUs and GPGPUs are also infeasible: they were not designed for I/O at 10-100Gbps; they also have asymmetric resources for input and output and thus cannot be pipelined[4, 2], whereas custom chip-based solutions are costly and inflexible to changes, and FPGA-based solutions are historically hard to program[6]; (2) There is a distinct advantage to utilizing high-bandwidth or line-speed analytics to reduce time-to-discovery of information, particularly ones that can be pipelined together to conduct a series of processing tasks or data tests without impeding data rates; (3) There is potentially significant network infrastructure cost savings possible from compact and power-efficient analytic support deployed at the network periphery on the data source or one hop away; (4) There is a need for agile deployment in response to changing objectives; (5) There is an opportunity to constrain designs to use only secure components to achieve their specific objectives. We address these five problems in our stream processor design to provide secure, easily specified processing for low-latency, low-power 10-100Gbps in-line processing on top of a commodity high-end FPGA-based hardware accelerator network processor. With a standard interface a user can snap together various filter blocks, like Legos™, to form a custom processing chain. The overall design is a four-layer solution in which the structurally lowest layer provides the vast computational power to process line-speed streaming packets, and the uppermost layer provides the agility to easily shape the system to the properties of a given application. Current work has focused on design of the two lowest layers, highlighted in the design detail in Figure 1. The two layers shown in Figure 1 are the embeddable portion of the design; these layers, operating at up to 100Gbps, capture both the low- and high frequency components of a signal or stream, analyze them directly, and pass the lower frequency components, residues to the all-software upper layers, Layers 3 and 4; they also optionally supply the data-reduced output up to Layers 3 and 4 for additional processing. Layer 1 is analogous to a systolic array of processors on which simple low-level functions or actions are chained in series[5]. Examples of tasks accomplished at the lowest layer are: (a) check to see if Field 3 of the packet is greater than 5, or (b) count the number of X.75 packets, or (c) select individual fields from data packets. Layer 1 provides the lowest latency, highest throughput processing, analysis and data reduction, formulating raw facts from the stream; Layer 2, also accelerated in hardware and running at full network line rate, combines selected facts from Layer 1, forming a first level of information kernels. Layer 2 is comprised of a number of combiners intended to integrate facts extracted from Layer 1 for presentation to Layer 3. Still resident in FPGA hardware and hardware-accelerated, a Layer 2 combiner is comprised of state logic and soft-core microprocessors. Layer 3 runs in software on a host machine, and is essentially the bridge to the embeddable hardware; this layer exposes an API for the consumption of information kernels to create events and manage state. The generated events and state are also made available to an additional software Layer 4, supplying an interface to traditional software-based systems. As shown in the design detail, network data transitions systolically through Layer 1, through a series of light-weight processing filters that extract and/or modify packet contents. All filters have a similar interface: streams enter from the left, exit the right, and relevant facts are passed upward to Layer 2. The output of the end of the chain in Layer 1 shown in the Figure 1 can be (a) left unconnected (for purely monitoring activities), (b) redirected into the network (for bent pipe operations), or (c) passed to another identical processor, for extended processing on a given stream (scalability).

It is widely accepted that wireless channels decorrelate fast over space, and half a wavelength is the key distance metric used in link signature (LS) for security assurance. However, we believe that this channel correlation model is questionable, and will lead to false sense of security. In this project, we focus on establishing correct modeling of channel correlation so as to facilitate proper guard zone designs for LS security in various wireless environments of interest.

We present an architecture for the Security Behavior Observatory (SBO), a client-server infrastructure designed to collect a wide array of data on user and computer behavior from hundreds of participants over several years. The SBO infrastructure had to be carefully designed to fulfill several requirements. First, the SBO must scale with the desired length, breadth, and depth of data collection. Second, we must take extraordinary care to ensure the security of the collected data, which will inevitably include intimate participant behavioral data. Third, the SBO must serve our research interests, which will inevitably change as collected data is analyzed and interpreted. This short paper summarizes some of our design and implementation benefits and discusses a few hurdles and trade-offs to consider when designing such a data collection system.

This paper presents a system named SPOT to achieve high accuracy and preemptive detection of attacks. We use security logs of real-incidents that occurred over a six-year period at National Center for Supercomputing Applications (NCSA) to evaluate SPOT. Our data consists of attacks that led directly to the target system being compromised, i.e., not detected in advance, either by the security analysts or by intrusion detection systems. Our approach can detect 75 percent of attacks as early as minutes to tens of hours before attack payloads are executed.

In this study, we present a control theoretic technique to model routing in wireless multihop networks. We model ad hoc wireless networks as stochastic dynamical systems where, as a base case, a centralized controller pre-computes optimal paths to the destination. The usefulness of this approach lies in the fact that it can help obtain bounds on reliability of end-to-end packet transmissions. We compare this approach with the reliability achieved by some of the widely used routing techniques in multihop networks.

A key question that arises in rigorous analysis of cyberphysical systems under attack involves establishing whether or not the attacked system deviates significantly from the ideal allowed behavior. This is the problem of deciding whether or not the ideal system is an abstraction of the attacked system. A quantitative variation of this question can capture how much the attacked system deviates from the ideal. Thus, algorithms for deciding abstraction relations can help measure the effect of attacks on cyberphysical systems and to develop attack detection strategies. In this paper, we present a decision procedure for proving that one nonlinear dynamical system is a quantitative abstraction of another. Directly computing the reach sets of these nonlinear systems are undecidable in general and reach set over-approximations do not give a direct way for proving abstraction. Our procedure uses (possibly inaccurate) numerical simulations and a model annotation to compute tight approximations of the observable behaviors of the system and then uses these approximations to decide on abstraction. We show that the procedure is sound and that it is guaranteed to terminate under reasonable robustness assumptions.

Moving Target Defense (MTD) can enhance the resilience of cyber systems against attacks. Although there have been many MTD techniques, there is no systematic understanding and quantitative characterization of the power of MTD. In this paper, we propose to use a cyber epidemic dynamics approach to characterize the power of MTD. We define and investigate two complementary measures that are applicable when the defender aims to deploy MTD to achieve a certain security goal. One measure emphasizes the maximum portion of time during which the system can afford to stay in an undesired configuration (or posture), without considering the cost of deploying MTD. The other measure emphasizes the minimum cost of deploying MTD, while accommodating that the system has to stay in an undesired configuration (or posture) for a given portion of time. Our analytic studies lead to algorithms for optimally deploying MTD.

Information system developers and administrators often overlook critical security requirements and best practices. This may be due to lack of tools and techniques that allow practitioners to tailor security knowledge to their particular context. In order to explore the impact of new security methods, we must improve our ability to study the impact of security tools and methods on software and system development. In this paper, we present early findings of an experiment to assess the extent to which the number and type of examples used in security training stimuli can impact security problem solving. To motivate this research, we formulate hypotheses from analogical transfer theory in psychology. The independent variables include number of problem surfaces and schemas, and the dependent variable is the answer accuracy. Our study results do not show a statistically significant difference in performance when the number and types of examples are varied. We discuss the limitations, threats to validity and opportunities for future studies in this area.

The relationship between accountability and identity in online life presents many interesting questions. Here, we first systematically survey the various (directed) relationships among principals, system identities (nyms) used by principals, and actions carried out by principals using those nyms. We also map these relationships to corresponding accountability-related properties from the literature. Because punishment is fundamental to accountability, we then focus on the relationship between punishment and the strength of the connection between principals and nyms. To study this particular relationship, we formulate a utility-theoretic framework that distinguishes between principals and the identities they may use to commit violations. In doing so, we argue that the analogue applicable to our setting of the well known concept of quasilinear utility is insufficiently rich to capture important properties such as reputation. We propose more general utilities with linear transfer that do seem suitable for this model. In our use of this framework, we define notions of "open" and "closed" systems. This distinction captures the degree to which system participants are required to be bound to their system identities as a condition of participating in the system. This allows us to study the relationship between the strength of identity binding and the accountability properties of a system.

Security features are often hardwired into software applications, making it difficult to adapt security responses to reflect changes in runtime context and new attacks. In prior work, we proposed the idea of architecture-based self-protection as a way of separating adaptation logic from application logic and providing a global perspective for reasoning about security adaptations in the context of other business goals. In this paper, we present an approach, based on this idea, for combating denial-of-service (DoS) attacks. Our approach allows DoS-related tactics to be composed into more sophisticated mitigation strategies that encapsulate possible responses to a security problem. Then, utility-based reasoning can be used to consider different business contexts and qualities. We describe how this approach forms the underpinnings of a scientific approach to self-protection, allowing us to reason about how to make the best choice of mitigation at runtime. Moreover, we also show how formal analysis can be used to determine whether the mitigations cover the range of conditions the system is likely to encounter, and the effect of mitigations on other quality attributes of the system. We evaluate the approach using the Rainbow self-adaptive framework and show how Rainbow chooses DoS mitigation tactics that are sensitive to different business contexts.