CS Research Symposium Presentations Index

Thank you to these students for presenting their research! *These students will be presenting a brief lightning talk of their research.

Jiahao Cai* | Advisor: Yonghwi Kwon | Research Area: Security
Presentation Title: EVIL MASTERMIND: Delivering Malicious Payloads via Innocent Actors
Abstract: We propose a new sophisticated server-side web attack, called EvilMind, that leverages benign websites to collect attack-triggering inputs and translates the collected inputs to malicious payloads via a complicated reverse-engineering resilient compiler. The proposed malicious payload compiler is fundamentally challenging to analyze and reverse-engineer as it translates any inputs to some outputs without errors and failures on any inputs. Hence, analyzing its execution does not help understand attack-triggering inputs. Moreover, the proposed attack is resilient to forensic investigations because it leverages non-malicious contents from benign websites, misleading the investigation (e.g., forensic analyst) and hindering identification of the real malicious actors behind the attack. We have developed a fully automated prototype, MindGen, that can generate malware instances that can launch such attacks. 

Renqin Cai* | Advisor: Hongning Wang  | Research Area: AI
Presentation Title: Learning to Distill Dynamic Long-term Dependence for Sequential Recommendation
Abstract: Recommender Systems are becoming ubiquitous and greatly impact people’s behaviors, from recommending products in Amazon, to recommending queries in Google, and to recommending friends in Facebook, etc. The dynamics of users' interest over time pose challenges on making accurate recommendations. As users interact with different items sequentially on these online service platforms like Amazon or Google, the sequential order of such actions often reveals the dynamics of users' interest. Sequential recommendation, which considers the sequential order by modeling the dependence of future actions on historical actions concerning the distance among these actions, is proposed to improve the recommender systems. In this work, we propose to make use of the category information of the target action to retrieve relevant transition patterns for improving target action prediction. Specifically, we use a gating network to predict the category of the target action. With the predicted category, we then distill the long-range historical actions to capture long-term dependence with a long-term encoder. 

Henry Carscadden | Advisor: Ravi, Sekharipuram Subramaniam |  Research Area: AI 
Presentation Title: Forecasting Influenza by Relative Intensity
Abstract: This poster provides a summary of work performed this summer focused on time series forecasting. Areas investigated include ensemble methods, recurrent neural networks, tree methods, SVMs, and traditional time series models applied to influenza forecasts. 

Hanjie Chen* | Advisor: Yangfeng Ji  | Research Area: AI
Presentation Title: Building Hierarchical Interpretations in Natural Language via Feature Interaction Detection
Abstract: The interpretability of neural networks has become crucial for their applications in real-world with respect to reliability and trustworthiness. Existing explanation generation methods usually provide important features by scoring their individual contributions to the model prediction and ignore the interactions between features, which eventually provide a bag-of-words representation as an explanation. In natural language processing, this type of explanations is challenging for human users to understand the meaning of an explanation and draw the connection between explanation and model prediction, especially for long texts. In this work, we focus on detecting the interactions between features and propose a novel approach to build a hierarchy of explanations based on feature interactions. The proposed method is evaluated with three neural classifiers, LSTM, CNN, and BERT, on two benchmark text classification datasets. The generated explanations are assessed by both automatic evaluation measurements and human evaluators. Experiments show the effectiveness of the proposed method in providing explanations that are both faithful to models, and understandable to humans. 

Jiayi Chen* | Advisor: Aidong Zhang  | Research Area: AI
Presentation Title: A Graph-based Machine Learning Approach to Incomplete Multimodality Data Analysis
Abstract: In many application scenarios such biomedicine, affective computing, social media, robotic and human-robot interaction, data usually comes in different types, namely multi-modality. One problem of analyzing multimodal data is to fuse the complementary information in multiple data sources. Another key problem in multimodal data analysis is the incomplete data problem. Some modalities are more easily missing because of different reasons. We propose a graph-based machine learning method to deal with incomplete multimodal data problem, aiming to model and learn the complicated multimodal data interactions from incomplete multimodal data using the designed graph structure and learning methods. 

Zhendong Chu | Advisor: Hongning Wang | Research Area: AI
Presentation Title: Accounting for Temporal Dynamics in Document Streams
Abstract: Textual information, such as news articles, social media, and online forum discussions, often comes in a form of sequential text streams. Events happening in the real world trigger a set of articles talking about them or related events over a period of time. In the meanwhile, even one event is fading out, another related event could raise public attention. Hence, it is important to leverage the information about how topics influence each other over time to obtain a better understanding and modeling of document streams. In this paper, we explicitly model mutual influence among topics over time, with the purpose to better understand how events emerge, fade and inherit. We propose a temporal point process model, referred to as Correlated Temporal Topic Model (CoTT), to capture the temporal dynamics in a latent topic space. Our model allows for efficient online inference, scaling to continuous time document streams. Extensive experiments on real-world data reveal the effectiveness of our model in recovering meaningful temporal dependency structure among topics and documents. 

Mitch Gerrard* | Advisor: Matthew Dwyer | Research Area: Software Engineering
Presentation Title: Conditional Quantitative Analysis
Abstract: We present a novel program analysis framework---conditional quantitative analysis (CQA)---that combines evidence from different underlying analyses to compute bounds on failure probability within a program. CQA builds on the principle of conditional verification, which combines the strengths of multiple verification procedures by excluding the portion of state space one of them verified as safe from the search space of the others. This allows each application of a verification procedure to focus only on parts of the state space which have not yet been covered by any of others. Because most verification procedures are specialized (or optimized) on specific program features (e.g., via specific abstractions or memory models), composing their partial results can harvest the respective strengths on portions of a program exhibiting specific features, leaving simpler residual problems after each application. By focusing expensive quantitative techniques on small portions of the state space, CQA can drastically improve upon the state-of-the-art in quantitative analysis, both with respect to runtime and precision of failure probability bounds. 

Ashley Gao | Advisor: Jack Stankovic | Research Areas: AI; Cyber-Physical Systems
Presentation Title: Monitor and Manage Stress of Caregivers’ to Dementia Patients
Abstract:  Many machine learning algorithms have been made in recent years to classify emotions. These algorithms, ranging from the classic approach such as SVM to the more recent deep learning, are all trained and evaluated on publicly available datasets of emotional speech. However, due to the subjective nature of emotion, the task to classify them is challenging, as the state-of-the-art solutions fail to perform satisfactorily. Meanwhile, only training and evaluating on the datasets that are collected from the controlled lab or studio environments where additional acoustic events, such as background noise, are kept minimal. Even if the classifiers achieve near-zero error rate during their testing phase, it is unsound to claim that they will perform as well in a realistic scenario in which additional acoustic events are omnipresent.  To address the two disadvantages of the state-of-the-art classifiers, we propose a new architecture of three CNN-DNN hybrid classifiers that are trained on not only the datasets devoid of additional acoustic events but also synthetic data to which these acoustic events are deliberately fabricated into. Our evaluation of the architecture shows a near-zero error rate on the original datasets and outperforms the state-of-the-art on the synthetic dataset. 

Minbiao Han* | Advisor: Micheal Albert, Haifeng Xu | Research Areas: Theory; AI
Presentation Title: Optimal auctions with strict incentive compatibility
Abstract: Different mechanisms have been proposed to design revenue-maximizing auctions, and the major difficulty among them is to incentivize bidders to report their private valuations truthfully. While there has been some work that can achieve approximate optimal incentive compatible auctions with high probability, they do not offer any guarantees on the non-truthful valuations. In our work, we aim to design optimal auctions with strict incentive compatibility which can guarantee the revenue of optimal auctions even if bidders do not report their valuations truthfully. 

Bargav Jayaraman | Advisor: David Evans | Research Areas: Security; AI 
Presentation Title: Evaluating Differentially Private Machine Learning in Practice
Abstract: Differential privacy is a strong notion for privacy that can be used to prove formal guarantees, in terms of a privacy budget, ϵ, about how much information is leaked by a mechanism. When used in privacy-preserving machine learning, the goal is typically to limit what can be inferred from the model about individual training records. However, the calibration of the privacy budget is not well understood. Implementations of privacy-preserving machine learning often select large values of ϵ in order to get acceptable utility of the model, with little understanding of the impact of such choices on meaningful privacy. Moreover, in scenarios where iterative learning procedures are used, relaxed definitions of differential privacy are often used which appear to reduce the needed privacy budget but present poorly understood trade-offs between privacy and utility. In this paper, we quantify the impact of these choices on privacy in experiments with logistic regression and neural network models. Our main finding is that there is no way to obtain privacy for free – relaxed definitions of differential privacy that reduce the amount of noise needed to improve utility also increase the measured privacy leakage. Current mechanisms for differentially private machine learning rarely offer acceptable utility-privacy trade-offs for complex learning tasks: settings that provide limited accuracy loss provide little effective privacy, and settings that provide strong privacy result in useless models. 

Yiling Jia | Advisor: Hongning Wang  | Research Areas: Cyber-Physical Systems; AI
Presentation Title: A Tree-Structured Neural Network Model for Household Energy Breakdown 
Abstract: Residential buildings constitute roughly one-third of the total energy usage across the globe. Numerous studies have shown that providing an \emph{energy breakdown, i.e., per-appliance energy consumption}, increases residents' awareness of energy use and can help save up to 15\% energy. A significant amount of prior work has looked into source-separation techniques collectively called non-intrusive load monitoring (NILM), and most prior NILM research has leveraged high-frequency household aggregate data (1 minute or less sampling interval) for energy breakdown. However, in practice most smart meters only sample hourly or once every 15 minutes, and existing NILM techniques show poor performance at such a low sampling rate. In this paper, we propose a tree-structured convolutional neural network (CNN) for energy breakdown on low frequency data. There are three key insights behind the design of our model: i) households consume energy with regular temporal patterns (e.g. time of day, day of week, etc.), which can be well captured by filters learned in CNNs; ii) tree structure isolates the pattern learning of each appliance that helps avoid the magnitude variance problem, while preserves the relationship among appliances; iii) tree structure enables the separation of known appliance from unknown ones, which de-noises the input time series for better appliance-level reconstruction. Our TreeCNN model outperforms seven existing baselines on a publicly available dataset with lower estimation error and higher accuracy on detecting the active states of appliances. 

Mainuddin Ahmad Jonas | Advisor: David Evans | Research Areas: Security; AI
Presentation Title: Measuring and Reducing Joint Vulnerability of Models to Adversarial Examples
Abstract: In this poster, we describe our on-going project on measuring, training and certifying jointly robust ensembles against adversarial examples. 

Josephine Lamp* | Advisor: Lu Feng | Research Areas: Cyber-Physical Systems
Presentation Title: A Logic Based Learning Approach to Explore Diabetes Behaviors
Abstract: Type I Diabetes (T1D) is a chronic disease in which the body’s ability to synthesize insulin is destroyed. It can be difficult for patients to manage their T1D, as they must control a variety of be- havioral factors that affect glycemic control outcomes. In this paper, we explore T1D patient behaviors using a Signal Temporal Logic (STL) based learning approach. STL formulas learned from real patient data characterize behavior patterns that may result in varying glycemic con- trol. Such logical characterizations can provide feedback to clinicians and their patients about behavioral changes that patients may implement to improve T1D control. We present both individual- and population-level behavior patterns learned from a clinical dataset of 21 T1D patients. 

Jack Lanchantin | Advisor: Yanjun Qi | Research Areas: AI 
Presentation Title: Neural Message Passing for Multi-Label Classification
Abstract: Multi-label classification (MLC) is the task of assigning a set of target labels for a given sample. Modeling the combinatorial label interactions in MLC has been a long-haul challenge. We propose Label Message Passing (LaMP) Neural Networks to efficiently model the joint prediction of multiple labels. LaMP treats labels as nodes on a label-interaction graph and computes the hidden representation of each label node conditioned on the input using attention-based neural message passing. Attention enables LaMP to assign different importances to neighbor nodes per label, learning how labels interact (implicitly). The proposed models are simple, accurate, interpretable, structure-agnostic, and applicable for predicting dense labels since LaMP is incredibly parallelizable. We validate the benefits of LaMP on seven real-world MLC datasets, covering a broad spectrum of input/output types and outperforming the state-of-the-art results. Notably, LaMP enables intuitive interpretation of how classifying each label depends on the elements of a sample and at the same time rely on its interaction with other labels. 

Marzieh Lenjani*  | Advisor: Kevin Skadron  | Research Areas: Computer Systems
Presentation Title: Fulcrum: a Simplified Control and Access Mechanism toward Flexible and Practical in-situ Accelerators
Abstract: Many modern and emerging applications are memory intensive, where the cost of moving data from memory to processor dominates the cost of computations. In-situ approaches process data very close to the memory cells, in the row buffer of each subarray, immediately after the data are read out. This minimizes data movement costs and also affords parallelism across subarrays. However, current in-situ approaches are limited to only row-wide bitwise (or few-bit) operations applied uniformly across the row buffer. Therefore they cannot support common operations (32-bit addition and multiplications), operations with data dependency, and operations based on predicates. Moreover, with current peripheral logic, communication among subarrays is inefficient and bits in a word are not physically adjacent. Our proposed method, Fulcrum, addresses these issues. We propose a new lightweight access and control mechanism that sequentially processes the data, enables operations with data dependencies along the row buffer, and supports operations based on a predicate. The sequential control mechanism requires only a one-word, scalar ALU, at the edge of every two subarrays, capable of a range of common operations (bitwise, addition, and multiplication). We show that this one-word ALU outperforms a row-wide bitwise ALU. For algorithms that require communication among subarrays, we augment the peripheral logic with broadcasting capabilities and a previously-proposed method for low-cost inter-subarray data movement. In order to realize true subarray-level parallelism, we introduce a lightweight column-selection mechanism through shifting one-hot encoded values. This technique enables independent column selection in each subarray. Our evaluation shows that, on average, our method outperforms (i) a contemporary GPU (Nvidia P100 with HBM2) by 118×, and (ii) an ideal model of a GPU (which only accounts for the overhead of data movement) by 83×. 

Sihang Liu* | Advisor: Samira Khan  | Research Areas: Computer Systems; Software Engineering
Presentation Title: Redesigning the system stack for Persistent Memory
Abstract: The advancement of memory technologies has brought us a new type of memory, persistent memory (PM), that features the high-speed and byte-addressability of DRAM, and persistence and high-capacity of storage device.PM can be accessed directly by the processor through a load/store interface, bypassing the OS indirection, and thus blurring the boundary between memory and storage. This unification of memory and storage provides a new opportunity to achieve unprecedentedly high performance of manipulating persistent data, while requires significant change to the whole system stack, from the underlying hardware that manages PM device, all the way to the operating system and user applications that maintain persistent data. At the hardware-level, as PM is persistent, it requires security protection for user data and mitigation for write endurance; at the software-level, programs' direct management of persistent data lifts the bar of programming on PM, and thus requires toolchains to aid the testing of program's correctness. As such, there is a renewed urgency in redesigning the system for PM. Our recent works aim to mitigate issues across the system stack. Our work Janus mitigates the performance degradation due to security and endurance support for PM systems; and, PMTest aims to detect bugs in PM programs that can lead to an inability to recover and data loss in event of a system failure. Janus improves performance by 2x and PMTest has detected bugs in Intel's PM-optimized file system and customized user-space programs. 

Zetian Liu* | Advisor: Haiying Shen  | Research Areas: Computer Systems
Presentation Title: RL-based scheduling for ML jobs
Abstract: We propose a new task scheduler that minimizes the average job completion time by cutting off some less important tasks of machine learning jobs. Our method also makes refinement in reducing the communication between distant nodes and optimizing the order of awaiting jobs in the queue. 

Meiyi Ma | Advisor: John Stankovic, Lu Feng | Research Areas: Cyber-Physical Systems
Presentation Title: CityResolver: A Decision Support System for Conflict Resolution in Smart Cities
Abstract: Resolution of conflicts across services in smart cities is an important yet challenging problem. We presentCityResolver – a decision support system for conflict resolution in smart cities. CityResolver uses an Integer Linear Programming based method to generate a small set of resolution options, and a Signal Temporal Logic based verification approach to compute these resolution options’ impact on city performance.The trade-offs between resolution options are shown in a dashboard to support decision makers in selecting the best resolution. We demonstrate the effectiveness of CityResolver by comparing the performance with two baselines: a smart city without conflict resolution, and CityGuard which uses a priority rule-based conflict resolution. Experimental results show that CityResolver can reduce the number of requirement violations and improve city performance significantly. 

Saeed Mahloujifar* | Advisor: Mohammad Mahmoody |  Research Areas: Theory; Security; AI
Presentation Title: Universal Multi-Party Poisoning Attacks
Abstract: In this work, we demonstrate universal multiparty poisoning attacks that adapt and apply to any multi-party learning process with arbitrary interaction pattern between the parties. More generally, we introduce and study (k, p)-poisoning attacks in which an adversary controls k ∈ [m] of the parties, and for each corrupted party Pi , the adversary submits some poisoned data T'_i on behalf of P_i that is still “(1−p)-close” to the correct data T_i (e.g., 1 − p fraction of T'_i is still honestly generated). We prove that for any “bad” property B of the final trained hypothesis h (e.g., h failing on a particular test example or having “large” risk) that has an arbitrarily small constant probability of happening without the attack, there always is a (k, p)-poisoning attack that increases the probability of B from µ to µ^(1−p·k/m) = µ+Ω(p·k/m).Our attack only uses clean labels, and it is online, as it only knows the the data shared so far. 

Sergiu Mosanu* | Advisor: Mircea Stan and Kevin Skadron | Research Areas: Security
Presentation Title: Flexi-AES: A Highly-Parameterizable Cipher for a Wide Range of Design Constraints
Abstract: We need efficient and secure communication between interconnected devices even via untrusted networks. For that, we employ hardware accelerated end-to-end security protocols using various encryption algorithms. Existing state-of-the-art implementations were tedious to adapt to a wide set of design constraintsWe present an open-source, flexible, feature-rich hardware implementation of the Advanced Encryption Standard (AES) in Chisel, simple to employ to any architecture / design requirements. 

Joseph Outten | Advisor: Andrew Warren | Research Areas: AI 
Presentation Title: Protein Sequence Annotation using Convolutional Neural Networks
Abstract: Proteins are the building blocks of biological processes. They are made up of functional subunits called domains. These domains, including binding and catalytic sites, if known, can shed light on the function of a protein and inform categorizations into potential threat levels for applications in biosecurity, bioinformatics, and synthetic biology. Current methods for annotating proteins use large databases of hidden markov models (HMMs), one per domain. Our work aims to replace these HMMs with a set of convolutional neural networks to increase runtime and accuracy of protein annotation. 

Josh Priest | Advisor: Sekharipuram Ravi (ssr6nh), Madhav Marathe | Research Areas: Theory 
Presentation Title: Solving Synchronous Dynamical Systems Problems Using Boolean Satisfiability
Abstract: Synchronous Dynamical Systems (SyDS) are graphs which have states and functions assigned to each node. The state space for each node is Boolean (0 or 1). Each state is updated at the same time as every other node. Every node has a function that defines how its state will change based on its local neighbors (nodes it has connections to). By default, nodes count themselves as one of their local neighbors. SyDSs can be used to model power systems, traffic systems, the internet, and more. The Boolean Satisfiability Problem (abbreviated as SAT) is the problem of determining whether, given a Boolean formula F, if there is a True or False (1, 0) value that can be assigned to each variable in F such that F evaluates to True. If there is, the SAT problem is considered satisfiable, and if not then it is considered unsatisfiable. There are many open source SAT solvers that are typically very fast as solving these SAT problems. The goal of this research is to solve SyDS problems using these open source SAT solvers. Two main problems are worked with: First, whether there exists a configuration of a SyDS that is its own successor. This configuration would be called a fixed point, or a stable configuration. The other problem asks, given a trajectory of configurations with some unknown node states, if there is a way to assign those unknown node states Boolean values such that the trajectory represents a real trajectory of the SyDS. This problem is called the Configuration Sequence Completion (CSC) problem. SyDS problems can be converted to SAT and given to open source SAT solvers to be solved quickly. 

Tamjid Al Rahat* | Advisor: Yuan Tian | Research Areas: Security; Software Engineering 
Presentation Title: Mobile developers use OAuth APIs to implement Single-Sign-On services
Abstract: Mobile developers use OAuth APIs to implement Single-Sign-On services. However, the OAuth protocol was originally designed for the authorization for third-party websites not to authenticate users in third-party mobile apps. As a result, it is challenging for developers to correctly implement mobile OAuth securely. These vulnerabilities due to the misunderstanding of OAuth and inexperience of developers could lead to data leakage and account breach. In this paper, we perform an empirical study on the usage of OAuth APIs in Android applications and their security implications. In particular, we develop OAuthLint, that incorporates a query-driven static analysis to automatically checkprograms on the Google Play marketplace. OAuthLint takes as input an anti-protocol that encodes a vulnerable pattern extracted from the OAuth specifications and a program P. Our tool then generates a counter-example if the anti-protocol can match a trace of P’s possible executions. To evaluate the effectiveness of our approach, we perform a systematic study on 600+ popular apps which have more than 10 millions of downloads. The evaluation shows that 101 (32%) out of 316 applications that use OAuth APIs make at least one security mistake. 

Elaheh Sadredini* | Advisor: Kevin Skadron | Research Areas: Computer Systems 
Presentation Title: Emerging Algorithm/Architecture Solutions for Pattern Matching
Abstract: High-throughput and concurrent processing of thousands of patterns on each byte of an input stream is critical for many applications with real-time processing needs, such as network intrusion detection, spam filters, virus scanners, and many more. The demand for accelerated pattern matching has motivated several recent in-memory accelerator architectures for automata processing, which is an efficient computation model for pattern matching. Our key observations are: (1) all these architectures are based on 8-bit symbol processing (derived from ASCII), and our analysis on a large set of real-world automata benchmarks reveals that the 8-bit processing dramatically underutilizes hardware resources, and (2) multi-stride symbol processing, a major source of throughput growth, is not explored in the existing in-memory solutions. We propose algorithm/architecture solutions to overcome these issues. 

Prathyush Sambaturu* | Advisor: Anil Vullikanti | Research Areas: Theory; AI 
Presentation Title: Designing robust intervention strategies to control epidemic outbreaks
Abstract: We study the problem of controlling the spread of an epidemicoutbreak on a network by deploying vaccines. Typically, the supplyof vaccines is limited, and becomes available over time. Therefore, vaccinationdecisions have to be made in multiple stages. These are challengingoptimization problems when the networks are arbitrary. Further, manycomponents of the epidemic model, e.g., the source of the outbreak mightnot be known, and parameters might be not known precisely. In such asetting, a min-max objective, which minimizes the maximum outbreaksize in any scenario (which corresponds to specif c source or parametersetting), gives a more robust solution, than tailoring the intervention toa single scenario.In this paper, we develop rigorous approximation algorithms for vaccineallocation in single and two-stages, for the min-max objective. We evaluateour algorithms on different random graph models. We find that thenodes picked in the interventions at different stages have very differentcharacteristics.We also find that solutions optimized for a single scenariocan be significantly sub-optimal with respect to the min-max objective. 

Rasool Sharifi* | Advisor: Ashish Venkat | Research Areas: Computer Systems; Security
Presentation Title: CHEx86: Context-Sensitive Enforcement of Memory Safety via Microcode-Enabled Capabilities
Abstract: This work introduces the CHEx86 processor architecture for securing applications, including legacy binaries, against a wide array of security exploits that target temporal and spatial memory safety vulnerabilities such as out-of-bounds accesses, use-after-free, double-free, and uninitialized reads, by instrumenting the code at the microcode-level, completely under-the-hood, with only limited access to source-level symbol information. In addition, this work presents a novel scheme for speculatively tracking pointer arithmetic and pointer movement, including the detection of pointer aliases in memory, at the machine code-level using a configurable set of automatically constructed rules. This architecture outperforms the address sanitizer, a state-of-the-art software-based mitigation by 54%, while eliminating porting, deployment, and verification costs that are invariably associated with recompilation. 

David Shriver* | Advisor: Sebastian Elbaum | Research Areas: Software Engineering 
Presentation Title: Differencing Neural Networks
Abstract: Differencing Neural Networks. Deep neural networks (DNN) have dramatically increased in capability and application domains over recent years. Just like traditional systems, DNNs evolve during development and deployment, for example, to improve performance, accuracy, and robustness. Techniques for assessing differences across these DNNs, however, is largely limited to statistical performance measures which render no information about the DNN semantic differences that may underlie those measures. In this work, we introduce an approach to uncover behavioral differences for DNNs, that is, inputs that may cause two DNNs to have different behavior. The approach is unique in that it guides a generative adversarial network (GAN) to produce inputs that reveal differences and are relevant with respect to the training distribution. Our study shows the potential of the approach in uncovering such relevant differences over two families of networks. 

Anshuman Suri* | Advisor: David Evans | Research Areas: Computer Systems; Cyber-Physical Systems; Security; AI
Presentation Title: Adversarial Machine Learning 

Fnu Suya | Advisor: David Evans | Research Areas: AI 
Presentation Title: Hybrid Batch Attacks: Finding Black-box Adversarial Examples with Limited Queries
Abstract: Estimate the cost for a black-box adversary to find adversarial examples. 

Swapna Thorve* | Advisor: Madhav Marathe | Research Areas: Computer Systems; Cyber-Physical Systems; AI; Software Engineering
Presentation Title: Energy Demand Modeling for Residential Sector
Abstract: Residential energy demand dynamics at household level can be studied through demographic, behavioral and physical characteristics of the household. In this paper, we develop an agent-based model using a bottom-up approach to build dis-aggregated energy demand estimates at the household level at an hourly interval. A household level analysis is made possible via the use of synthetic populations for the urban and rural areas of Virginia, USA. The energy consumption estimate is based on householders' demographics, their behaviors and activities, ratings of appliances used in energy-related activities, space conditioning fuels, physical characteristics of the home, and weather conditions. Results from the simulation are then validated with actual demand curves from Rappahannock county in Virginia using dynamic time warping. The simulation results show that the model produces realistic energy demand profiles. 

Md Fazlay Rabbi Masum Billah, Nabeel Nasir, Nurani Saoda, Alexander Sarris, Wenpeng Wang, Brad Campbell | Advisor: Brad Campbell 
Research Areas: Computer Systems; Cyber-Physical Systems
Abstract: The accelerating progress towards millimeter-scale computers, increasing aversions to replacing batteries, and growing contention in unlicensed wireless spectrum all contribute to constrain embedded devices, yet the need for sophisticated applications, concerns over user privacy, and demonstrated potential of machine learning all demand more from emerging IoT devices. We present our group's work on relieving this tension through architectures and applications for device-to-device applications running entirely at the edge, new abstractions for energy-harvesting sensors, and miniaturized machine learning models running on microcontrollers. These new platforms help keep sensor data streams local, help developers write applications with batteryless devices, and expand the inference capability of in-situ sensors. 

Yizhou Wei* | Advisor: Samira Khan | Research Areas: Computer Systems
Presentation Title: PIMProf: A Performance Profiler for Processing-in-Memory Architectures
Abstract: PIMProf: A Performance Profiler for Processing-in-Memory Architectures. PIM architectures have drawn an increasing research interest as a mitigation to the data movement bottleneck within the current DRAM-based architectures, and a variety of them have been proposed for accelerating various data-intensive workloads. However, for a given workload, it is difficult to determine which part of a program should be offloaded to a given PIM architecture to gain the best performance and how much performance gain is possible. We propose PIMProf, a tool that uses a combination of static and runtime analysis to automatically detect the PIM candidates and estimate the speedup of the program. 

Trent Weiss* | Advisor: Madhur Behl  | Research Areas: Cyber-Physical Systems; AI
Presentation Title: DeepRacing
Abstract: We consider the challenging problem of high speed autonomous racing in realistic dynamic environments. DeepRacing is a novel end-to-end framework, and a virtual testbed for training and evaluating algorithms for autonomous racing. The virtual testbed is implemented using the realistic Formula One (F1) Codemasters game, which is used by many F1 drivers for training.