Understanding potential attacks against privacy systems using advanced machine learning

Computer Science Department Seminar

Speaker:  Se Eun Oh, Ewha Womans University

Date and Time: October 11, 2024  9-10am MDT

Location: CCSB1.410 and Teams

 

Abstract: 

Tor is one of the most widely-used anonymous network systems with millions of daily users. It supports a layered encryption using three hops to route the traffic between client and server, which in turn hides the both ends of the communication link. In this talk, we introduce two fundamental attacks on Tor, website fingerprinting and end-to-end flow correlation, which  analyze network traffic metadata such as traffic volume and timing information. These styles of attacks successfully enable potential adversaries to infer users’ online activities and deanonymize the communication ends, which alarms Tor researchers to design appropriate defenses guaranteeing the privacy of Tor users.

Biography:

Host:

Mohammad Saidur Rahman, msrahman3@utep.edu

Understanding the Privacy Dimension of Wearables through Machine Learning-enabled Inferences

Computer Science Department Seminar

Speaker:  David Mohaisen, University of Central Florida

Date and Time: September 20, 2024   11am-12noon

Location: CCSB 1.410 and Teams

 

Abstract: 

To keep up with the ever-growing user expectations, developers keep adding new features and to augment the use cases of wearables, such as fitness trackers, augmented reality head mounted devices (AR HMDs), and smart watches, without considering their security and privacy impacts. In this talk, I will introduce our recent results on understanding the privacy dimension of wearables through inference attacks facilitated by machine learning.

To start, I will present an exploration of the attack surface introduced by fitness trackers where we propose an inference attack that breaches location privacy through the elevation profiles collected by fitness trackers. Our attack highlights that adversaries can infer the location from elevation profiles collected via fitness trackers. Then, I will review the attack surface introduced by smartwatches by developing an inference attack that exploits the smartwatch microphone to capture the acoustic emanations of physical keyboards and successfully infers what the user has been typing. Finally, I will present an exploration of the AR HMD’s through the design of inference attack that exploits the geometric projection of hand movements in air. The attack framework predicts the typed text on an in-air tapping keyboard, which is only visible to the user.

Biography:

David Mohaisen (PhD’12, University of Minnesota) is a Full Professor of Computer Science at the University of Central Florida, where he has been since 2017. Previously, he was an Assistant Professor at SUNY Buffalo (2015-2017) and a Senior Scientist at Verisign Labs (2012-2015). His research interests are in the broad area of applied security and privacy, covering aspects of networked systems, software systems, IoT and AR/VR, machine learning, and blockchain systems. His research has been supported by a number of generous grants from NSF, NRF, AFRL, AFOSR, etc., and has been published in top conferences and journals alike, with multiple best paper awards. His work was featured in multiple outlets, including the New Scientist, MIT Technology Review, ACM Tech News, Science Daily, etc. Among other services, he has been an Associate Editor of various journals, including IEEE TDSC, IEEE TMC, and IEEE TPDS.  He is a senior member of ACM (2018) and IEEE (2015), a Distinguished Speaker of the ACM and Distinguished Visitor of the IEEE Computer Society. More information about his work is at  https://www.cs.ucf.edu/~mohaisen/

Host:

Mohammad Saidur Rahman, msrahman3@utep.edu

FlipPath game to counter stealthy attacks in sdn-based tactical networks

fabrice mvah

When: friday, may 10, 2024 from 10:30 - 11:30 am

 

Abstract:

Software-Defined Networking (SDN) is an evolving network paradigm that separates the control and forwarding functions of network devices. Its application in tactical networks can help to automate virtual private network establishments, unit task reorganization, security certificate management, and security solution automation. Despite these advantages, tactical networks continue to face new threats. In this work, we introduce a new stealthy packet-dropping scenario that can appear during data forwarding in tactical networks. To overcome this attack scenario, we propose a new game-theoretic approach called FlipPath to provide an optimal data forwarding strategy that prevents packet-dropping on a network path. The forwarding path is considered a shared resource, and each player wants to control it for a long period while minimizing the related costs. We use periodic strategies to characterize the Nash equilibria that identify the defender's optimal data-forwarding strategies. As the defender plays periodically, we use cyber deception to motivate the attacker to maintain its periodic strategy by making it believe that its attacks were successful. Our computational results have shown that the defender who adopts the proposed methodology can mitigate the attacker's stealthy packet-dropping.

Bio:

Fabrice Mvah is a Ph.D. candidate at the Department of Mathematics and Computer Science of University of Dschang, Cameroon. His doctoral research is funded by a scholarship from the US Army Research Laboratory under ARL-Africa, an initiative to increase research collaboration between the US Army Research Laboratory and universities in Africa. He has co-authored 5 scientific papers in top tier journals and conferences such as GameSec, EEE ICNC, COSE, J Hardw Syst Secur, and Int. J. Inf. Secur. He received his B.S. degree in Mathematics and Computer Science from University of Dschang in 2016, and the M.S. degree in Network and Distributed Services from the same University in 2018. His research interests include cybersecurity, game theory, software-defined network and cyber deception.

 

Uncertainty Aspects in Digital Twins

Niklas Winnewisser

Institute for Risk and Reliability, Leibniz University Hannover

When: Friday, March 8, 2024 from 11am-12pm

Where: CCSB 1.0410

 

Abstract:

Decision-making that enables managing engineering systems to be more resilient and cost-optimized is paramount to navigating the complexity of today's increasingly uncertain environmental and political landscape. Tackling this challenge immerses us in the realm of digital twins and demands a broad perspective on the incorporated modeling environments. If we think of a system as a potentially dynamic collaboration of components working towards a common objective – e.g., providing sufficient network efficiency – then the example of a transportation infrastructure network represents a system of roads and critical engineering structures. We consider critical components such as bridges to be complex systems in their own right.

At the heart of a digital twin of an engineering system lies the consolidation of diverse information onto a unified platform, encompassing data from sensors and experts, structural response models for safety and degradation assessments, and Building Information Modeling for construction and maintenance logistics. Here, uncertainties are ubiquitous. In conventional analytical frameworks, uncertainties in models arise from variations, e.g., in material properties and loading conditions, as well as from inaccuracies, e.g., in measurements, prediction models, and expert assessments. In addition, the integration of linked data allows us to assign quantifiable indicators for the information quality of statements and the credibility of their sources within and along the boundaries of the digital twin. This possibility implies exploring how this metadata space can enable self-monitoring and automated improvements or adjustments to measurement, modeling, and action strategies.

We start with a broad overview of multiscale digital twins and dive into the first fundamental insights of engineering approaches. Afterward, we will see current research questions related to clustering, optimization, decision-making, and information quality management.  

 

Host: Vladik Kreinovich, vladik@utep.edu

 

 

Perceptions of Pragmatic Similarity amoing Spoken Utterances: A Model and its Applications

Nigel Ward, Andy Segura, Divette Marco

CS Department Seminar

When: February 16, 11am - 12pm

Where: BUSN 312

 

Abstract:

Screening for autism and  other communication disordedrs in early childhood can improve patient outcomes, and we wish to use AI to help. The obvious approach, using supervised machine learning methods to learn to classify children, is ruled out by data limitations, so we are exploring an alternative, knowledge-rich method. The strategy is to use general model of perceived pragmatic similarity, able to take two utterances and output an estimate of how similarly they will be perceived. To create this model, we first collected 5496 human judgements fo the similarity between utterance pairs. We then built various models, of which the best -- cosine similarity over 103 features selected from a deep pretrained model, HuBert -- correlates as highly as 0.74 with the average human judgement. This model enabled good performance, 82% accuracy, on the task of classifying adolescents as neurotypical or autistic, using k nearest neighbors. This model also supports diverse use cases in the diagnosis and treatment workflow, which we will illustrate with a live demo. 

Acknowledgment:

The work was supported in part by Meta and by NSF's AI Institute for Exceptional Education.

 

The CENATE Project A Research Sample

Joseph Manzano

Pacific Northwest National Laboratory

When: October 20, 11:00 am - 12:00 pm, 

Where: BUSN 302

Abstract

The Center for Advanced Technology Evaluation (CENATE) is an evaluation effort for novel and cutting-edge hardware advancing scientific computing, AI/ML, and cybersecurity. The center provides a platform for computational scientists, computer scientists, and industrial technology developers to exchange and test new ideas, collaborations, and technologies while aligning themselves with the Department of Energy's scientific missions. To present the research developed in CENATE, we have selected three research directions in the area of networks: 1) creating boutique application-specific topologies, 2) using reinforcement learning to optimize wireless communication parameters, and 3) improving cybersecurity by detecting denial of service attacks using a novel methodology. 

Bio

Joseph B. Manzano is a senior computer scientist in the High-Performance Computing group at the Pacific Northwest National Laboratory. He was born in Guayaquil, Ecuador, and started his academic career in New Jersey, USA in 1999. He received his Ph.D. from the University of Delaware in 2011 and has been a Pacific Northwest National Laboratory computer scientist since then. His research career has touched on topics like dataflow runtimes and architectures, system software for multi-core and distributed systems, performance evaluation and modeling of extreme-scale systems, cybersecurity for high-performance systems, and distributed machine learning systems. He is also part of the organizing committees for several high-performance conferences and workshops on a large set of topics, including power for extreme and novel systems (E2SC series of workshops), system software for high-performance computers (IPDRM workshop series), compiler technologies for codesign in ML/AI workflows (MLH/S workshop series at the PACT conference), cybersecurity in supercomputers (S-HPC workshops at SC 22 and 23), among others. He has also participated in several technical committees for conferences, journals, and PhD defenses. Joseph has published more than 60 peer-reviewed publications between conference, workshop, and journal papers

Hosts: Shirley Moore and Deepak Tosh.

Computer arithmetic for DNN acceleration or how to compute right with errors and do it fast 

Anastasia Volkova,

Université de Nantes 

When: September 1, 11:00 am - 12:00 pm, 

Where: BUSN 302

Abstract

Numerical algorithms rarely use exact arithmetic to perform their computations; instead they employ more efficient floating-point or fixed-point representations. The use of these finite-precision numerical formats results in computational errors that influence the result. On the other hand, the choice of precision influences the performance (latency, memory usage) of the implemented algorithm. 

In the search of a sweet spot between the accuracy and the efficiency, we first establish the relation between the precision and accuracy for a given algorithm and then optimise arithmetic parameters (data formats, function approximations, hardware operators) such that accuracy requirements are satisfied at minimal cost. Ideally, this process should be automatic and applicable to large-scale systems.

In this talk we showcase a range of tools and techniques for arithmetic optimisations for DNN inference. We act on three levels:

• Automatic analysis of numerical quality of a DNN model and establishing the minimal accuracy requirement for each layer of a network; 
• Building custom hardware operators for dot-product and employing the power of Integer Linear Programming to optimise the resources; 
• Retraining DNNs in a hardware-friendly manner. 

Bio

Anastasia Volkova is an Associate Professor in Computer Science at the University of Nantes, France. She conducts her research at the LS2N lab in team OGRE. Previously, she was a research resident at Intel Corporation in San Diego, USA and a postdoctoral researcher at INRIA in Lyon, France. She is a former PhD student of Christoph Lauter and Thibault Hilaire at Sorbonne University UPMC in Paris, France. 

Host: Christoph Lauter 

 

Development of a Digital Twin-based Diabetes Decision Support System

Gyorgy Eigner

When: April 28, 11:00 am - 12:00 pm, 

Where: BUSN 332

Abstract

The diabetes diary, diagnostic and therapeutic toolkit being developed under the auspices of the Physiological Controls SystemResearch Center of Óbuda University will be presented. Examples of such developments include AI-powered blood glucose prediction and closed-loop control systems will be introduced. Furthermore, the beta version of these developments and the University of Óbuda's cloud services framework will be presented. 

Bio

GYORGY EIGNER (GSM’13–M’16–SM’20) got his B.Sc. degree in Mechatronical Engineering at Obuda University, Bánki Donát Faculty of Mechanical and Safety Engineering in 2011, and M.Sc. degree in Biomedical Engineering at Budapest University of Technology and Economics in 2013. He received his Ph.D. degree at Obuda University in 2017, where he is currently associate professor and the Dean of John von Neumann Faculty of Informatics. His main research focus is the application of advanced control methods in physiological relations. On these areas he published more than 100 scientific work. He is the professional leader of the Zsambek Future Industries Science Park. He is the Co-Chair of the Computational Cybernetics Technical Committee, Director of the Robotics Special College of the Obuda University. 

 

Performance Prediction Toolkit (PPT): A Scalable Performance Modeling and Prediction Approach for GPUs and CPUs

Professor Hameed Badawy

Host: Dr. Moore

When: Friday, March 3, 11:00 A.M. - 12:00 P.M.

Where: CCSB 1.0202

 

Abstract

With the advent of Machine Learning (ML) in many fields, highly efficient computational resources are necessary. GPUs top the charts for such ML-oriented computational resources. Measuring the performance of GPUs provides insights into what is achievable. We will present PPT focusing on PPT-GPU, a scalable performance prediction toolkit for GPUs. PPT-GPU achieves scalability through a hybrid high-level modeling approach where some computations are extrapolated, and multiple parts of the model are parallelized. The tool’s primary prediction models use pre-collected memory and instructions traces of the workloads to capture the dynamic behavior of the kernels accurately.

PPT-GPU reports an extensive array of GPU performance metrics accurately while being easily extensible. We use a broad set of benchmarks, including DeepBench, an ML workload, to verify the prediction accuracy. We compare the results against hardware metrics collected using vendor profiling tools and cycle-accurate simulators. The results show that the performance predictions are highly correlated to the actual hardware (MAPE: < 16\% and Correlation: > 0.98) on NVIDIA’s V100 GPUs. Moreover, PPT-GPU is orders of magnitude faster than cycle-accurate simulators. This comprehensiveness of the collected metrics can guide architects and developers to perform design space explorations. Furthermore, the tool’s scalability enables efficient and fast sensitivity analyses for performance-critical applications. We will show some results on NVIDIA’s A100. We will quickly touch on our work on PPT modeling of CPUs

Bio

Prof. Abdel-Hameed (Hameed) A. Badawy (Senior Member, IEEE) received a B.Sc. degree (Hons.) in Electronics Engineering from Mansoura University, Egypt, with a concentration on computers and control systems. He obtained his M.Sc. and Ph.D. in computer engineering from the University of Maryland. He is a tenured associate professor with the Klipsch School of Electrical and Computer Engineering, New Mexico State University, Las Cruces, NM, USA. He is the Computer Area Chair in the department. Also, he is a joint faculty with the Los Alamos (LANL) and Berkeley (LBNL) labs. He has been a visiting research scientist with the New Mexico Consortium. He was a lead research scientist with the High-Performance Computing Laboratory (HPCL) at George Washington University. His research interests include Performance Modeling, Prediction, Monitoring, and Evaluation, High-Performance Computer Architecture, Post-Moore’s law hardware, Hardware Security including IoT devices, Machine Learning for Computational problems, and Quantum Computing.

 

 

TACOS and FAJITA para todos!

Marcelo Frias, Professor at Instituto Tecnológico de Buenos Aires

Host: Dr. Vladik Kreinovich, vladik@utep.edu, Dr. Maria Mariani, mcmariani@utep.edu​

When: Tuesday, October 18, 1:30-2:30pm

Where: LART 205

Abstract

In this talk I will present the current version of the bug-finding tool TACO. TACO, for Translation of Annotated COde, allows one to automatically detect if a program annotated with a contract complies with, or violates, said contract. In case a contract violation is detected, a program input exposing the violation is automatically synthesized by TACO.

Along the talk I will use the tool and explain the technical foundations that make TACO highly effective to detect failures in pogroms over dynamically allocated data structures (a particularly difficult domain). Also, I will mention applications of TACO in automated test input generation through the FAJITA tool, and automated program repair.

Intended Audience

While the talk should be of particular interest for researchers in Software Engineering, I believe undergrad students are the ones that will be most impressed and are specially welcomed.

Bio

Professor at Instituto Tecnológico de Buenos Aires
Program Committee member of FM’19 (Formal Methods 2019), International Conference on Formal Methods, Porto, Portugal, 7-11 October, 2019.
Program Committee Member, Automated Software Engineering, ASE 2019.
Program Committee Member, International Conference on Software Testing, ICST 2017.
Program Committee Member, Automated Software Engineering, ASE 2016.
Program Committee Member, International Conference on Software Test- ing, ICST 2016.
Program Board Member, International Conference on Software Engineering, ICSE 2016.

 

 

An introduction to the Aberdeen Architecture: High Assurance Hardware State Machine Microprocessor Concept

Patrick Jungwirth, Ph.D., Computer Engineer, DEVCOM Army Research Laboratory

Host: Jaime Acosta, jcacosta@utep.edu

When: Friday, October 14th, 11:00 A.M - 12:00 P.M.

(A Teams link for the virtual talk will be sent prior to the talk)

Where: CCSB G.0208

Abstract

In a traditional computer, an operating system manages computer system resources. Current microprocessors execute or run instructions without any verification or authentication. There is no difference between safe instructions, coding errors, and malicious instructions. Complete mediation is a computer security principle meaning to verify access rights and authority for every operation. The Aberdeen Architecture enforces Saltzer and Schroeder’s security principles down to the instruction execution level and it provides near to full complete mediation for instruction execution.

The Aberdeen Architecture is also designed to block information leakage. It uses state machines to monitor and manage the execution pipeline. The state machine monitors completely virtualize the execution pipeline.  State machine monitors manage and track information flows:  (1) instruction execution flow, (2) control flow, (3) data flow, and (4) memory access flow.  Each information flow class has an allowed set of operations.  Security properties of data are completely tracked and managed from data creation (data source) until data is retired (data sink)

 

Bio

Patrick Jungwirth is currently a computer architecture researcher at the Army Research Lab. He is also an adjunct professor teaching computer architecture courses at the University of Maryland, Baltimore County. He previously worked at the Aviation and Missile, Research Development and Engineering Center. He holds two computer architecture patents with one pending. His research interests cover computer architectures and digital signal processing. He has published a textbook on sampling theory and analog-to-digital conversion.

 

Industry 4.0 and Smart Cities

Miroslav Svítek

Host: Vladik Kreinovich, vladik@utep.edu

When: Thursday, October 6th, 1:30 P.M. -  2:50 P.M.

Where: Lart 305

Abstract

Main characteristics of Industry 4.0 include horizontal integration through networks in order to facilitate an internal cooperation, vertical integration of subsystems within the factory in order to create a flexible and adaptable manufacturing systems. Another basic feature of Industry 4.0 is a Cyber-Physical System (CPS) as a system of collaborating computational elements controlling physical entities up to creation Digital Twin of the factory.

The Smart City approach, developed at Czech Technical University in Prague, is based on these principles of Industry 4.0, which are further researched and practically applied in “Smart Europská Avenue” that plays the role of a digital testbed or living laboratory of Prague city.

Smart Evropská Avanue uses a variety of sensors, starting with physical detectors, cameras, and ending with space imaging (weather prediction, city temperature maps, and emission maps). It should be noted that even a vehicle or a mobile phone in this concept becomes an intelligent sensor providing important data. The city management, thanks to current data, moves from the original predefined dynamic plans to adaptive control algorithms that ensure the coordination of entire territorial units. Different simulation tools are used to validate individual strategies. In virtual space, it is much easier to model responses to different types of extraordinary events.

 

Bio

Prof. Miroslav Svítek graduated in radioelectronic from Czech Technical University in Prague, in 1992. In 1996, he received the Ph.D. degree in radioelectronic at Faculty of Electrical Engineering, Czech Technical University in Prague. Since 2008, he has been Full professor in engineering informatics at Faculty of Transportation Sciences, Czech Technical University in Prague. In 2010 – 2018 he was Dean of Faculty of Transportation Sciences, Czech Technical University in Prague. Since 2018, he has been Visiting professor in smart cities at University of Texas at El Paso, USA. The focus of his research includes complex system sciences and their practical applications to Intelligent Transport Systems, Smart Cities and Smart Regions. He is the author or co-author of more than 200 scientific papers and 10 books.

 

Early Exascale Hardware and Applications

Shirley Moore, University of Texas at El Paso

Host: Shirley Moore

When: Friday, October 7th, 11:00 A.M

Where: CRBL 205

Abstract

An exascale computer is capable of executing at least 10 ¹⁸ double precision floating point operations per second. The world’s first exascale computer is Frontier at Oak Ridge National Laboratory. Frontier was announced in June 2022 as the world’s fastest supercomputer on the TOP500 list. This achievement is the result of several years of collaboration between government agencies, lab researchers, and hardware vendors on the design of three exascale machines: Frontier, Aurora, and El Capitan. I will describe the process that was used to evaluate and select from the candidate exascale architectures and then give details of the selected architectures. Finally, I will describe three exascale application efforts I have been involved with in the areas of fusion plasma modeling, simulation of heterogeneous catalysis using computational chemistry, and simulation of cancer cell transport through the human circulatory system.

 

Bio

Shirley Moore is an Associate Professor in the Computer Science Department at the University of Texas at El Paso. Prior to returning to UTEP in fall of 2020, she was a Senior Computer Scientist in the Future Technologies Group at Oak Ridge National Laboratory (ORNL) for four years, where she led ORNL participation in several Exascale Computing Project (ECP) efforts. She received her PhD in Computer Sciences from Purdue University in May of 1990 in the area of distributed database systems. Her research interests are in parallel and distributed computing, including edge computing, with a focus on performance evaluation and optimization.

 

Programmatic Reinforcement Learning for All

Ashutosh Trivedi, University of Colorado Boulder

Host: Saeid Tizpaz Niari

When: Friday, September 30th, 11am-12pm

Where: CCSB G.0208

Abstract

Reinforcement Learning (RL) is an optimization-based approach to problem-solving where learning agents rely on scalar reward signals to discover optimal solutions. The recent success of RL has demonstrated its potential as a viable alternative to "human" programming. However, observing these success stories closely, it is evident that significant expertise is required in deploying RL. This expertise is required in designing a suitable approximation architecture and designing a suitable "flat" representation of the environment in a form required by the architecture. Besides, it is also needed to specify objectives in the language of scalar rewards. This rigid interface---in the form of feature constructions, manual approximations, and reward engineering---between the programmers and the RL algorithms is cumbersome and error-prone. The resulting lack of usability and trust contributes towards barriers to entry in this promising field.  My group is working towards democratizing reinforcement learning by developing principled methodologies and powerful tools to improve the usability and trustworthiness of RL-based programming at scale.

The aforementioned low-level interactions between the programmers and the RL are akin to programming systems in a low-level assembly language. I envision a programmatic approach to RL where the programmers interact with the RL algorithms by writing programs in a high-level programming language expressing the simulation environment, the choices available to the learning agent, and the learning objectives, while an underlying “interpreter” frees the programmer from the burden of feature construction and approximation heuristics demanded by the state-of-the-art RL algorithms. We dub this setting high-level programmatic reinforcement learning (or programmatic RL for short).

To realize the promise of improved usability of programmatic RL, we need RL algorithms capable of efficiently handling rich programmatic features (functional recursion and recursive data structures) and complex dynamical models (governed by ordinary differential equations) while guaranteeing convergence to the optimal value.  To enable transparent and trustworthy RL, we need translation schemes to compile learning requirements expressed in high-level languages to scalar reward signals. In this talk, I will summarize our efforts and breakthroughs towards a framework for programmatic RL capable of reasoning with formal requirements, real-time constraints, and recursive environments.

 

Bio

Ashutosh Trivedi received his B.Eng. in Computer Science from NIT Nagpur in 2000, his M.Tech. in Electrical Engineering from IIT Bombay in 2003, and his Ph.D. in Computer Science from the University of Warwick in 2009. He was a postdoctoral researcher at the University of Oxford and at the University of Pennsylvania between 2009 and 2012. He was an assistant professor of Computer Science at IIT Bombay from 2013-2015. He joined the University of Colorado Boulder in 2015, where he is currently an assistant professor of Computer Science. He is also a member of the Programming Languages and Verification (CUPLV) Group. His research interests include formal methods, optimization, and game theory with applications in trustworthy AI, cyber-physical systems, software security, and fairness in AI. He is a recipient of an NSF CAREER award, two AFRL fellowships, and a Liverpool-India fellowship.

 

Internal eigenvalues using constraint interval theory with applications

Marina Tuyako Mizukoshi, UFGo

Host: Professor Kreinovich

When: Friday, August 5, 11:00 A.M - 12:00 P.M

Where: CCSB G.0208

Abstract

This talk will introduce the constraint interval theory (CI).  The CI will be applied to study interval eigenvalues in interval symmetric matrix. The interval eigenvalue is important to obtain conditions about stability in a dynamical systems with uncertainty in their coefficients.

 

Professor Mizukoshi, of the Institute of Mathematics and Statistics at the Universidade Federal de Goias in Brazil, is currently Visiting Professor at the University of Colorado, Denver.

Computational Intelligence For Engineering Solutions: Invariance-Based Approach
(a brief overview of the Fall 2022 class CS 5354/CS 4365)

Dr. Kreinovich

When: Friday, April 22, 2:30pm - 3:30pm 

Where: Business, room 318

Abstract

The main purpose of computers is to process real-life data, so that we will able to understand the current state of the system and to predict its future behavior. In some situations -- e.g., in basic mechanics -- we have fundamental from-first-principles laws that enable us to make the corresponding predictions. However, in many other situations,

especially in engineering, we only have approximate empirical formulas. For example, it is not possible to predict, based on the

first principles,  how pavement will deteriorate with time, or how people will change their opinions about goods.

 

In such situations, we face the following problems:

• Why these formulas? Users are usually reluctant to use purely empirical formulas. Reason: there is no guarantee that these

              formulas will work in a new situations. It is therefore desirable to come up with theoretical explanations for these formulas.

• Maybe these formulas are not the best. Within these theoretical explanations, are the current formulas most adequate? And if they are not the best, what are the better formulas?
• What next?  Empirical formulas are usually approximate. If we want a more accurate description, we need more complex, more

              detailed formulas. Of course, the ultimate test is comparing with the observations and measurement results. In view of the

              theoretical explanations, what are good candidates for such more complex formulas?

Similarly, in many engineering applications, there are semi-empirical methods for solving the corresponding problems. In such cases, similar problems appear.

 

In this course, on the example of engineering and computer science applications, we will explain how invariance-based techniques – widely used in physics – can help explain and improve empirical formulas and methods.

 

See https://www.cs.utep.edu/vladik/cs5354.22/syllabus.html for a detailed description of the class

Socio-Cognitive Factors of Decision making in Spear-Phishing Attacks

Dr. Prashanth Rajivan

When: Friday, April 15, 11am - 12pm 

Where: Virtual

Abstract

Despite significant advancements in security technologies, phishing attacks continue to be rampant and successful because it is cognitively challenging for humans to distinguish phishing emails from real messages. One phishing email and one vulnerable person is all it takes for an attacker to succeed. To combat the rampant phishing threats, companies rely primarily on machine learning algorithms for automated detection, and on the human ability to detect attacks that algorithms miss. Although current algorithms are successful in detecting known mass-phishing messages, they do not guarantee complete protection.

In this talk, I will discuss experiments we are conducting to understand human decision-making process in the context of phishing. First, I will describe a new simulation paradigm we have developed for studying human behavior in spear-phishing attacks from both the attacker and end-user perspective. Next, I will present results from a cognitive model based on Instance Based Learning Theory developed to predict and analyze human response to phishing emails obtained from a laboratory experiment. I will describe the effectiveness of integrating natural language processing methods, such as, GloVe, and BERT with cognitive models to predict human response to phishing emails. Finally, I will introduce follow-on research directions I am currently pursuing in phishing and other related areas such misinformation.

Bio

Prashanth Rajivan is an assistant professor of Industrial and Systems Engineering and adjunct assistant professor of human centered design and engineering at the University of Washington. His research agenda is on the intersection of human factors and computer security. His areas of interests include security and privacy decision making, simulation and modeling, computer supported cooperative work, and applied cognitive science. Prior to this appointment, Prashanth Rajivan was a Postdoctoral Research Fellow at the Department of Social and Decision Sciences, Carnegie Mellon University, Pittsburgh. He holds a Ph.D. in Human Systems Engineering (2014) and M.S. in Computer Science (2011) from Arizona State University, USA. He is the author of several peer-reviewed publications and book chapters. His work on multi-agent models of teamwork in cyber defense was awarded the best student paper at HFES annual conference in 2014. His dissertation work was a finalist in the Human Factors Prize on Cyber Security in 2017. His research is funded by NSF (CAREER), AHRQ, Starbucks and CISCO.

 

Host

Dr. Palvi Aggarwal 

 

CS Teaching Innovations Talk

Dr. Diego Aguirre; Dr. Eric Freudenthal

When: Friday, April 1, 11am - 12pm 

Where: CCSB G.0208

Talk 1

In this short talk, I will present strategies I have explored in my attempt to cultivate a growth mindset among CS 2302 (Data Structures) students. In particular, I will present: 1) alternative grading systems that reward growth without sacrificing rigorousness, 2) reflection-based assignments that reinforce deliberate practice tenets,  and 3) sample course assignments and activities designed to enhance students’ competencies in key problem-solving skills (e.g., problem decomposition, abstraction, solution analysis and prioritization).

Talk 2

What Happened when I (Sorta) Let Students set the Rules.  The pandemic hit our students badly and the extension of conventional classroom policies to online seemed punitive.  Instead, I asked students what they wanted and (mostly) gave it to them.   After a bit of adjustment and reality checks, this strategy seems to have improved both teaching teams’ and student experiences.  We’re back in the classroom, and my pandemic-motivated policies are (mostly) working out fine.

 

Fairness-aware Configuration of Machine Learning Libraries

Dr. Saeid Tizpaz-Niari

When: Friday, February 18^th 2022, 11:00 AM

Where: Business 331

Abstract

This paper investigates the parameter space of machine learning (ML) algorithms in aggravating or mitigating fairness bugs. Data-driven software is increasingly applied in social-critical applications where ensuring fairness is of paramount importance. The existing approaches focus on addressing fairness bugs by either modifying the input dataset or modifying the learning algorithms. On the other hand, the selection of hyperparameters, which provide finer controls of ML algorithms, may enable a less intrusive approach to influence the fairness. Can hyperparameters amplify or suppress discrimination present in the input dataset? How can we help programmers in detecting, understanding, and exploiting the role of hyperparameters to improve the fairness?

We design three search-based software testing algorithms to uncover the precision-fairness frontier of the hyperparameter space. We complement these algorithms with statistical debugging to explain the role of these parameters in improving fairness. We implement the proposed approaches in the tool Parfait-ML (PARameter FAIrness Testing for ML Libraries) and show its effectiveness and utility over five mature ML algorithms as used in six social-critical applications. In these applications, our approach successfully identified hyperparameters that significantly improve (vis-a-vis the state-of-the-art techniques) the fairness without sacrificing precision. Surprisingly, for some algorithms (e.g., random forest), our approach showed that certain configuration of hyperparameters (e.g., restricting the search space of attributes) can amplify biases across applications. Upon further investigation, we found intuitive explanations of these phenomena, and the results corroborate similar observations from the literature.

Bio

The key theme in Dr. Tizpaz-Niari  research is to automate the process of finding and explaining bugs, vulnerabilities, and fairness issues in large-scale software and machine learning systems. In particular, he is interested in analyzing a differential class of properties that include confidentiality, privacy, and fairness. The findings help discover multiple performance bugs in popular ML libraries such as scikit-learn, fairness bugs in the state-of-the-art machine learning algorithms, and timing side-channel vulnerabilities in critical Java libraries such as OpenJDK and Apache ([1],[2],[3],[4],[5]). I lead the Responsible, Informative, and Secure Computing Lab (RISC) at UTEP.

 

Resolving the Privacy Paradox: Towards Trustworthy and Collaborative AI

Rui Hu

When: Friday, February 4^th 2022, 10:00 – 11:00 AM

Where: CCSB 1.0410

Abstract

As an immense number of connected devices such as mobile devices, wearables, and autonomous vehicles generate massive amounts of data each day to develop machine learning (ML) based intelligent services, multiple spheres of human life, such as healthcare, entertainment, and industry are being transformed. The traditional process for developing machine learning applications is to gather a large dataset, train a model on the data, and run the trained model on a cloud server. Due to the growing tension between the need for big data and the need for privacy protection, it is increasingly attractive to enable edge devices to collaboratively train ML models while keeping the data locally. However, the deployment of this collaborative ML architecture depends on a set of challenges, such as the new privacy risks, limited resources, heterogeneous data and devices, and security vulnerabilities. In this talk, I will cover the privacy-preserving collaborative ML approaches that address the new privacy risks in collaborative ML. I will also discuss my observations and responses to other key challenges, as well as important research directions, with respect to the development of collaborative ML.

Bio

Rui Hu received the B.Eng. degree in Electrical Engineering from Jinan University, China, in 2017. She is currently a Ph.D. candidate in the Department of Electrical and Computer Engineering at the University of Texas at San Antonio. Her research interest includes security and privacy, machine learning, edge computing, and Internet-of-Things. She was awarded the Dorrough Distinguished Graduate Fellowship and Graduate Student Professional Development Awards in 2018 and 2019, respectively.

 

Robots, Language, and Representations

Dr. Nakul Gopalan

When: Tuesday, February 1^st, 2022, 3:00 – 4:00 PM

Where: Prospect Hall, Room 234 

Abstract

Robots are increasingly present in our lives, from cleaning our houses to automating logistics. However, these robots are still present in our lives as solitary agents, performing structured tasks, without the power to collaborate and learn with humans. 

 A key challenge here is that robots perceive the world and operate in it using sensors and actuators that are continuous, low-level and noisy. However, people on the other hand, reason, plan, specify and teach tasks using high-level concepts without worrying about the low-level continuous nature of the world. To address this challenge, I develop computational methods that firstly, allow robots to learn representations, and skills to solve novel tasks. Moreover, these methods, and representations also enable robots to be taught and programmed using natural language communication, allowing robots to understand a human partner's intent.

 In this talk I first demonstrate how representations for planning and language understanding can be learned together to follow commands in novel environments. In the second part of the talk, I demonstrate a more practical approach in which language can be grounded to pre-trained deep policy representations to solve novel task specifications.

Together, these approaches empower robots to learn unstructured tasks via language and demonstrations. I will then discuss the implications of such approaches in collaborative task solving with robots in homes, offices and industries.

Bio

Nakul Gopalan is a postdoctoral researcher in the CORE Robotics Lab with Prof. Matthew Gombolay at Georgia Tech. He completed his PhD at Brown University's Computer Science department in 2019. Previously he was a graduate student in Prof. Stefanie Tellex's H2R lab at Brown. His research interests lie at the intersection of language grounding and robot learning. Nakul has developed algorithms and methods that allow robots to be trained by leveraging demonstrations and natural language descriptions. Such learning would improve the usability of robots within homes and offices. His other research interests are in hierarchical reinforcement learning and planning. His work has received a best paper award at the RoboNLP workshop at ACL 2017.

 

Modeling the 3D Human Genome Structure

Dr. Oluwatosin Oluwadare

When:Thursday, January 27^th, 2022, 10:00-11:00 am 

Where: Prospect Hall, Room 234

Abstract

Nineteen years after the sequencing of the human genome and 20 years after the introduction of Chromosome Conformation Capture (3C) technologies, three-dimensional (3D) inference and big data remain problematic in the field of genomics, and specifically, in the 3C data analysis research area. Chromosome 3D structure inference involves reconstructing a genome’s 3D structure or, in some cases, an ensemble of structures from contact interaction frequencies extracted from Hi-C experiments. Further questions remain about chromosome topology and structure; enhancer promoter interactions; location of genes, gene clusters, and transcription factors; the relationship between gene expression and epigenetics; and chromosome visualization at a larger scale. Oluwatosin Oluwadare’s research interests lie at the intersection of computer science— particularly artificial intelligence (AI) and data mining—and biological studies—particularly genomics. With a focus on developing machine learning and data mining methods to analyze big biomedical data and address fundamental problems in biomedical sciences. In this presentation, he will discuss his research about human genome 3D structure reconstruction from high throughput sequencing data (Hi-C). Also, he will present the algorithms “3DMax” and “CBCR” for chromosome and genome 3D structure prediction from Hi-C data. The talk will also highlight “GenomeFlow,” a comprehensive GUI tool to facilitate the entire modeling, analysis, and visualization of the 3D genome organization.

Bio

Dr. Oluwatosin Oluwadare is an Assistant Professor of Computer Science and Innovation at the University of Colorado, Colorado Springs (UCCS). He received his Bachelor of Technology degree in Computer Science (CS) from the Federal University of Technology, Akure, Nigeria, his Master of Science degree in CS from the University of Texas, Arlington, and his Ph.D. in CS from the University of Missouri, Columbia. Dr. Oluwadare’s research focus areas are Bioinformatics & Computational Biology, Machine Learning, Deep Learning, and Big data analytics. He has proposed and been published in reputable journals and has developed novel methods focused on machine learning applications in bioinformatics. Dr. Oluwadare is the Director of the Oluwadare Lab research group at UCCS, where he and his students focus on developing computational methods to address complex biological questions. More details about his research and research group can be found at https://academics.uccs.edu/~ooluwada/.

Dr. Oluwadare has a keen interest in researching in machine learning and its various applications. He proposed and led the development of a software app called EyeCYou; this app uses AI to provide the facial description of a person to the visually impaired. Learn More about EyeCYou here: http://eyecyouapp.com/. The app is freely available for download on the apple app store.

 

Towards Reliable Machine Learning

When: Friday, January 21st, 2022, 10:00-11:00 am

Where: Prospect Hall, Room 234

Abstract

Deep Neural Networks (DNN) represent a performance-hungry application. Floating-Point (FP) and custom floating-point-like arithmetic satisfies this hunger. While there is need for speed, inference in DNNs does not seem to have any need for precision. Many papers experimentally observe that DNNs can successfully run at almost ridiculously low precision. Whenever fast FP is absent, such as on embedded controllers, Fixed-Point (FxP) Arithmetic gets drawn in. FxP however has its own particular challenges.

 The aim of this talk is many-fold: first, to shed some theoretical light upon why a DNN's FP accuracy stays high for low FP precision. We observe that the loss of relative accuracy in the convolutional steps is recovered by the activation layers, which are extremely well-conditioned.  We give an interpretation for the link between precision and accuracy in DNNs.

 Second, in this talk, we would like to give an overview on the principles of error analysis for FP and FxP arithmetic. We shall derive the basic machine epsilon from the IEEE754 FP definitions and demonstrate its use in different error models. We shall introduce interval arithmetic as a tool for computing without "forgetting" about error. We shall also comment on what challenges stem from the historical way IEEE754 is written and what we see as possible solutions these historical hindrances.

 Third, the talk presents a software framework for semi-automatic FP and FxP error analysis for the inference phase of deep-learning. Compatible with common Tensorflow/Keras models, it leverages the frugally-deep Python/C++ library to transform a neural network into C++ code in order to analyze the network's need for precision. This rigorous analysis is based an Interval and Affine arithmetics to compute absolute and relative error bounds for a DNN. The tool is able to determine the Most Significant Bit (MSB) position of every FxP variable in a given DNN model, in order to ensure that no overflow can occur while maintaining appropriate output accuracy. While word-length is currently still fixed, this is feature of our tool is unique. We demonstrate our tool with several examples.

 Finally, we shall provide an outlook on what we think is still missing for completely Reliable Machine Learning.  

Bio

 Christoph Lauter was born in Germany. He received his BS and MS degrees in Computer Science (Minor in Electrical Engineering) from Technische Universität München, Munich, in 2005. He received this Ph.D. in Theoretical Computer Science from École Normale Supérieure de Lyon, France, in 2008. From 2008 to 2010, he worked as a Software Engineer in the Numerics Team at Intel Corporation, in Hillsboro. Since 2010, he has been Maître de Conférences (Associate Professor) at Sorbonne Université, Paris, France. He has been on leave from this position since 2018, when he joined University of Alaska Anchorage (UAA), where he has held an Assistant Professor position. On leave from UAA, since January 2022, Christoph Lauter has rejoined Intel Corporation, in an attempt to bridge Academia and Industry, aiming at building a US School of Floating-Point Arithmetic.  His research interests are Floating-Point Arithmetic, Validated Numerical Computing and Reliable Machine Learning. He is a member of the IEEE754 Standard's Revision Committee and he was a ballotting member for the IEEE754-2019 revision.

 

Computer Vision for Advancing 3D Immersive Applications

Dr. Kevin Desai

When: Wednesday, December 8, 2021, 3:00-4:00pm

Where: Prospect Hall, Room 234

Abstract

Deep Neural Networks (DNN) represent a performance-hungry application. Floating-Point (FP) and custom floating-point-like arithmetic satisfies this hunger. While there is need for speed, inference in DNNs does not seem to have any need for precision. Many papers experimentally observe that DNNs can successfully run at almost ridiculously low precision. Whenever fast FP is absent, such as on embedded controllers, Fixed-Point (FxP) Arithmetic gets drawn in. FxP however has its own particular challenges. The aim of this talk is many-fold: first, to shed some theoretical light upon why a DNN's FP accuracy stays high for low FP precision. We observe that the loss of relative accuracy in the convolutional steps is recovered by the activation layers, which are extremely well-conditioned. We give an interpretation for the link between precision and accuracy in DNNs. Second, in this talk, we would like to give an overview on the principles of error analysis for FP and FxP arithmetic. We shall derive the basic machine epsilon from the IEEE754 FP definitions and demonstrate its use in different error models. We shall introduce interval arithmetic as a tool for computing without "forgetting" about error. We shall also comment on what challenges stem from the historical way IEEE754 is written and what we see as possible solutions these historical hindrances. Third, the talk presents a software framework for semi-automatic FP and FxP error analysis for the inference phase of deep-learning. Compatible with common Tensorflow/Keras models, it leverages the frugally-deep Python/C++ library to transform a neural network into C++ code in order to analyze the network's need for precision. This rigorous analysis is based an Interval and Affine arithmetics to compute absolute and relative error bounds for a DNN. The tool is able to determine the Most Significant Bit (MSB) position of every FxP variable in a given DNN model, in order to ensure that no overflow can occur while maintaining appropriate output accuracy. While word-length is currently still fixed, this is feature of our tool is unique. We demonstrate our tool with several examples. Finally, we shall provide an outlook on what we think is still missing for completely Reliable Machine Learning.

 

Bio

Kevin Desai is an Assistant Professor of Instruction in the Computer Science department at the University of Texas at San Antonio. He received his PhD degree in Computer Science from The University of Texas at Dallas in May 2019 with his dissertation titled – “Quantifying Experience and Task Performance in 3D Serious Games”. He also received his MS in Computer Science from The University of Texas at Dallas in May 2015, whereas his Bachelor of Technology in Computer Engineering from Nirma University (India) in June 2013. Dr. Desai’s research experience and interests are in the fields of Computer Vision and Immersive (Virtual / Augmented / Mixed) Realities with applications in the domains of healthcare, rehabilitation, virtual training, and serious gaming. He conducts interdisciplinary research which mainly revolves around the real-time capture and generation of 3D human models and their incorporation in collaborative 3D immersive environments. Dr. Desai’s work has been published in peer-reviewed international conferences in the fields of computer vision (e.g., ICIP), VR / AR / MR (e.g., ISMAR), and Multimedia (e.g., MMSys, ISM, BigMM, ICME). He also serves as a program committee member and reviewer for top-tier international journals and conferences in IEEE, ACM, and Springer.

 

A Dimensional Model of Interaction Style Variation in Spoken Dialog

Dr. Ward, Professor, UTEP CS Department

Time: Friday, November 19th 2:30-3:30pm 

Location: CCSB 1.0202

Abstract

In spoken dialog people vary their interaction styles, and dialog systems should also be able to do the same. Previous work has elucidated many aspects of style variation and adaptation, but a general model has been lacking. We applied Principal Component Analysis over 84 novel features that encode the frequencies of diverse interaction-related prosodic behaviors.  The top 8 resulting dimensions together explained most of the variance, and were each interpretable.  Some dimensions represented well-known aspects of style (degree of engagement, talkativeness, factuality, etc.) and others novel ones (positive vs negative, change-oriented vs accepting, etc.).  Further, surprisingly, we found that individuals generally exhibited a wide variety of styles: using individual style tendencies to predict behavior outperformed an speaker-independent model by only 3.6%.

This talk reports work presented at SigDial 2021, plus new results from work with Jonathan E. Avila.

Predictive Intelligence to Enable Continuous Software Re-engineering and Preventive Maintenance 

Dr. Badreddin, Associate Professor, UTEP CS Department 

When: Friday, November 12th, 2:30-3:30pm 

Where: CCSB 1.0202 

Short Abstract

By predicting how software codebases evolve over time, software engineers can address maintenance and reliability issues before they manifest. The talk presents a recent collaborative NSF proposal along with early results. The talk covers broad topics, including data mining, machine learning, predictive analysis, and software engineering.

 

Long Abstract 

A key challenge in designing long-living software systems is the unpredictable nature of the software evolution. A sound design deemed effective today may result in a rapidly deteriorating unsustainable codebase over time. Reengineering activities that aim to address existing deficiencies may contribute to quality and reliability degradation in the long term. This challenge is particularly evident in experimental and research software systems. Scientists developing research software often follow domain-specific and knowledge-driven discovery processes. Research software is also subject to unique budgetary constraints and schedules. This, among other factors, further increase the unpredictability of the software evolution and can complicate its maintenance and sustainability.

If it is possible to predict quality and reliability issues early, then engineers can design software that avoids or minimizes the impact of these yet-to-come reliability and quality issues. Engineers can continuously re-design software as evolutionary scenarios become more probable. Moreover, it becomes possible to proactively address a reliability or maintenance issue even before it materializes.

Early results show that basic machine learning algorithms are highly effective in predicting key code quality and reliability metrics. The talk will highlight these results, and will present a recent NSF proposal submitted in collaboration with Dr. Hossain.

---

---

---

---

---

---

---

---

---