REU Site Summer Research Program

• Dr. Nihat Altiparmak

  

  Faculty Profile

  View Dr. Nihat Altiparmak’s faculty profile

  Research Areas:

  • Operating systems
  • Data storage systems
  • Parallel and distributed systems

  Sample Project Description:

  Energy-Aware Optimization of Operating System Kernels for Ultra-Low Latency Storage Performance

  The operating system kernel is a complex piece of software that sits between the applications of a computer and its hardware, with the main goal of providing convenient access to computer hardware in an efficient and fair manner. From personal computers to servers and supercomputers, embedded systems and robots to mobile devices like phones and tablets, operating systems are heavily used to manage a variety of computer systems in today’s world.

  Since computer hardware is constantly improving and innovating, operating system kernels need continuous optimizations in order to maintain their efficiency. Today’s operating systems urgently require optimization in their storage stack, which manages access to data storage devices. Various innovations have taken place in the storage subsystem within the past few years, including wide adoption of the Non-Volatile Memory Express (NVMe) interface [1] and the emergence of new generation of Ultra-Low Latency (ULL) SSDs, which are broadly defined as providing sub-10 μs data access latency [2]. This new level of ULL storage device performance questions the suitability of existing kernel storage stack designs that are primarily optimized for older storage generations with higher data access latencies, such as flash-based SSDs and HDDs, with one and three orders of magnitude slower data access latencies, respectively.

  Efficiency of an operating system is generally measured in terms of its performance, where the energy impact of the optimizations is commonly neglected. In other words, energy efficiency is typically regarded as a second class citizen in operating system design. However, today’s data centers consume as much electricity as a city [3]. In addition, the energy efficiency of an operating system managing a battery-operated device such as a mobile phone or tablet can significantly affect the battery life of that device. Therefore, any optimization that is performed on an operating system kernel should be performed in an energy-aware manner.

  In this REU project, the undergraduate student(s) will investigate, analyze, and optimize the Linux kernel’s storage stack for ULL storage performance in an energy-aware manner. Each student will be assigned a specific storage stack section to investigate, such as submission, scheduling, or completion, and perform empirical experiments using real hardware: an Intel Optane SSD [4] (a type of ULL SSD) in a dedicated server with an Onset HOBO Power meter [5] to measure energy consumption. Using this experimental setup, students will first investigate and understand the working principles of their assigned storage stack layer. Next, they will experimentally analyze the efficiency of existing methodologies within their layers in an energy-aware manner, and make observations about their deficiencies. Finally, based on their observations, students are expected to propose and implement possible optimizations to eliminate these observed deficiencies.  Students will be trained on kernel development and will work in collaboration with other lab students experienced in operating system design and development.

  References Cited

  • NVM Express Base Specification, rev. 1.4. https://nvmexpress.org/wp-content/uploads/NVM-Express-1_4-2019.06.10-Ratified.pdf.
  • Bryan Harris and Nihat Altiparmak. Ultra-Low Latency SSDs’ Impact on Overall Energy Efficiency. In 12th USENIX Workshop on Hot Topics in Storage and File Systems (HotStorage 20). USENIX Association, July 2020.
  • Bryan Harris and Nihat Altiparmak. Monte Carlo Based Server Consolidation for Energy Efficient Cloud Data Centers. In 11th IEEE International Conference on Cloud Computing Technology and Science (CloudCom 2019), pages 263–270, Sydney, Australia, December 2019.
  • Intel Optane SSD 900P Series Product Brief. https://www.intel.com/content/dam/www/public/us/en/documents/product-briefs/optane-ssd-900p-brief.pdf.
  • Onset HOBO® UX120-018 Data Logger (datasheet). Onset. https://www.onsetcomp.com/datasheet/UX120-018.

• Dr. Sabur Baidya

  

  Faculty Profile

  View Dr. Sabur Baidya’s faculty profile

  Research Areas:

  • Machine Learning
  • Edge/Fog Computing
  • Computer Systems

  Sample Project Description:

  Compressed Neural Networks for Multiscale Computing in Resource-constrained Systems

  Modern intelligent and autonomous systems are equipped with many sensors which are processed on single-board embedded computing platform. These systems often are battery powered with finite energy budget and they communicate with other devices for data/computation sharing over wireless networks. The constraint in computing, communication and energy, has led to the development of multiscale computing where data or computation can be opportunistically offloaded to an edge/fog server or to a cloud server [1]. However, if communication resources are insufficient then offloading data or commutation in raw form might not work and it necessitates compression of the data and/or computing algorithm. Additionally, with the boom of the Internet-of-things (IoT), devices are becoming lightweight which led to the development of TinyML [2] framework to support machine learning algorithms in small embedded boards and even on microcontrollers. However, depending on the application needs, the resource availability in the system and the constraints of available machine learning framework, the deep learning algorithms need to be compressed and/or optimally compiled to optimize the quality-of-service (QoS) of the applications.

  In this REU project, the undergraduate student will work on several Deep Neural Networks based applications for various sensor processing and investigate the following:

  • How can the models be compressed with some of the state-of the-art techniques, e.g., quantization [3], pruning [4], distillation [5], early-exit [6]. This will need studying the research papers to understand the theory behind the techniques and also implantation of the available open-source code for those techniques with Tensorflow or Pytorch on general purpose computer. This will need some knowledge of Machine learning, Python coding skills, and some experience of training ML models with Tensorflow or Pytorch.
  • Optimal compilation of the compressed models on different embedded platforms, e.g., Microcontrollers, Raspberry Pi, Nvidia Jetson Nano etc, with supported ML frameworks e.g., TVM, TF Lite, TF Lite Micro etc. The task will involve converting the ML flow graph to the supported ML framework for the corresponding platform. This does not need deep embedded system knowledge. Some experience on working with Unix/Linux based system will be good.
  • Create Pareto frontier with the accuracy, latency and energy performance tradeoff for various regime of computing resource availability on various platforms. This task will need measuring latency and accuracy of the algorithm, and energy consumption of the system.
  • Depending on availability of time, the student may try investigating the aforementioned characterizations for a few applications, e.g., Object Detection from image data, 3D object detection based on Lidar data etc.

  References Cited:

  1. Yousefpour, A., Fung, C., Nguyen, T., Kadiyala, K., Jalali, F., Niakanlahiji, A., Kong, J. and Jue, J.P., 2019. All one needs to know about fog computing and related edge computing paradigms: A complete survey. Journal of Systems Architecture, 98, pp.289-330.
  2. Warden, P. and Situnayake, D., 2019. Tinyml: Machine learning with tensorflow lite on arduino and ultra-low-power microcontrollers. O’Reilly Media.
  3. Hubara, I., Courbariaux, M., Soudry, D., El-Yaniv, R. and Bengio, Y., 2017. Quantized neural networks: Training neural networks with low precision weights and activations. The Journal of Machine Learning Research, 18(1), pp.6869-6898.
  4. Han, S., Pool, J., Tran, J. and Dally, W.J., 2015. Learning both weights and connections for efficient neural networks. arXiv preprint arXiv:1506.02626.
  5. Matsubara, Yoshitomo, Sabur Baidya, Davide Callegaro, Marco Levorato, and Sameer Singh. “Distilled split deep neural networks for edge-assisted real-time systems.” In Proceedings of the 2019 Workshop on Hot Topics in Video Analytics and Intelligent Edges, pp. 21-26. 2019.
  6. Teerapittayanon, S., McDanel, B. and Kung, H.T., 2016, December. Branchynet: Fast inference via early exiting from deep neural networks. In 2016 23rd International Conference on Pattern Recognition (ICPR)(pp. 2464-2469). IEEE.

• Dr. Adrian Lauf

  

  Faculty Profile

  View Dr. Adrian Lauf’s faculty profile

  Research Areas:

  • Embedded systems
  • Accelerators

  Sample Project Description:

  Embedded Systems and Accelerators

  Embedded systems and microcontrollers are so ubiquitous in their application breadth that virtually all categories of devices and components seem to carry them. From disposable conveniences to durable goods and control systems, microcontrollers and embedded-class devices are everywhere. Thanks to significant efforts by microcontroller core designers such as Espressif, Atmel/Microchip, and ARM, anyone with a minimal amount of programming skill can implement basic system programming architectures with easy-to-use and friendly Integrated Development Environments (IDEs), some of which can even be run in a browser, short-circuiting traditional frustrations of importing libraries and dependencies [1]. However, not all IDEs and basic implementations are created equal, especially with respect to understanding power consumption and sleep states. As embedded devices are produced by the billions each year [2], even infinitesimally-minute differences can scale to produce enormous consequences when considering the collective carbon footprint of these devices. In respect to this, the student will:

  1. Write a set of programs that feature IO-based, timer/counter-based, and watchdog timer-based wakeup events for a microcontroller that runs a process control system. The process control inputs and outputs will be simulated by software and transmitted to the MCU via serial interfaces. Two control systems will be present – an industrial automation system (e.g., a sorting system or automated assembly plant), and a power generation station. The MCU implementation will represent one node in the process.
  2. These programs will be written in the top 3-4 prototyping/programming languages, including C, Wiring (Arduino), and MicroPython.
  3. Power consumption for each of the sleep state programs will be carefully monitored by directly cutting power feeds to the microcontroller and inserting a high-precision ammeter with a capture interface, provided to the student.
  4. The student will then extrapolate the power consumption differences between the language choices and IDEs to understand how these choices in aggregate can contribute to the global carbon footprint and power consumption based on the billions of devices produced annually.

  Additionally, these microcontroller architectures now feature fully-integrated communications modules, also easily integrated into an IDE through manufacturer-provided libraries. Examples include Espressif ESP 8266 and ESP32-based microcontroller implementations that feature WiFi, Bluetooth, BLE, and LoRa radios [3-5]. It is not uncommon for prototyping boards and even full-fledged devices to feature multiple radio types. It is also increasingly common that device-to-device mesh networks can provide some relief to infrastructure-based networks by distributing communications distances and loads between nearby devices [6]. With reference to carbon footprint and energy consumption, the student will be provided with an Espressif ESP32-based development board with Bluetooth, BLE, WiFi, and LoRa radios to perform the following tasks:

  1. A simulated process control network will be provided to the student complete with nodes intended for monitoring, control, and actuation. A set of 7 of these devices will be provided in various physical configuration spaces.
  2. The student will leverage the various communications interfaces based on data rate, power consumption, and security to create a multiple-radio network to increase reliability and range, and to reduce power consumption. Furthermore, sleep modes that enable/disable the radios as needed will be integrated to save additional power.
  3. Power consumption with the multi-radio implementation will be studied against a baseline of heterogeneous radio networks (e.g., all nodes use WiFi) to see what gains can be made given the network performance requirements.
  4. The information gathered from this set of experiments will be extrapolated to include process control networks around the world to understand the carbon and energy usage impact of embedded network choices in microcontroller applications.

  This work will assist in developing metrics and guidelines that can have a large cumulative impact by generating an energy consumption mapping of how microcontroller devices contribute to energy consumption based on configuration and radio usage.

  References Cited

  1. M. Banzi and M. Shiloh, Getting started with Arduino: the open source electronics prototyping platform. Maker Media, Inc., 2014.
  2. R. Lineback, “Microcontrollers Will Regain Growth After 2019 Slump ” IC Insights.
  3. Espressif SYstems, “ESP32 Datasheet,” Iot based microcontroller, 2017.
  4. U. S. Z. Abidin, “Design and development of multi-transceiver Lorafi board consisting LoRa and ESP8266-Wifi communication module,” 2018.
  5. C. Bouras, A. Gkamas, V. Kokkinos, and N. Papachristos, “Using LoRa technology for IoT monitoring systems,” in 2019 10th International Conference on Networks of the Future (NoF), 2019: IEEE, pp. 134-137.
  6. F. Adelantado, X. Vilajosana, P. Tuset-Peiro, B. Martinez, J. Melia-Segui, and T. Watteyne, “Understanding the limits of LoRaWAN,” IEEE Communications magazine, vol. 55, no. 9, pp. 34-40, 2017.

• Dr. Olfa Nasraoui

  

  Faculty Profile

  View Dr. Olfa Nasraoui’s faculty profile

  Research Areas:

  • Machine learning
  • Fairness in AI

  Sample Project Description:

  Explainable Machine Learning Algorithms

  The most accurate Big Data predictive methods tend to rely on black-box machine learning models, such as Deep Learning, which lack interpretability and do not provide a straightforward explanation for their outputs. Yet explanations can improve the transparency of a predictive system by justifying predictions, and this in turn can enhance the user’s trust in the system. Hence, one main challenge in designing a machine learning model is mitigating the trade-off between an explainable technique with moderate prediction accuracy and a more accurate technique with no explainable predictions.

  

  Explanation mechanisms play a critical role in building trust into human-machine learning algorithm interaction and can therefore contribute to building ‘fair’ machine learning algorithms.  Fair algorithms make predictions that are not biased against any group of people. In general most unfair models are the result of biased data. The rationale for ensuring fairness through explanations is that an unfair or biased prediction would be easier to detect when an explanation is provided along with the prediction.

  In this project, an REU student will investigate a set of white box and black box machine learning predictive models and methods to build in explanation ability and fairness in black box methods.  Experimental results should demonstrate that the developed method is effective in generating accurate and explainable predictions.

• Dr. Rui Yu

  

  Faculty Profile

  View Dr. Rui Yu’s faculty profile

  Research Areas:

  • Computer vision
  • Machine learning
  • Human-computer interaction

  Sample Project Description:

  Enhancing Privacy in Live Camera Streams on Mobile Devices

  When people engage in live video streaming via mobile devices, their privacy can become compromised as the video may inadvertently reveal personal information. This issue is particularly pertinent for people with visual impairments (PVI). Remote sighted assistance (RSA) has emerged as a valuable assistive technology for PVI, wherein remote sighted agents offer real-time support via video-chat-like communication. However, prevalent RSA applications such as Aira [1] and Be My Eyes [2] have yet to implement robust measures to safeguard user privacy during live video sessions.

  This project aims to develop an iOS mobile app designed to selectively conceal sensitive information captured by the device’s camera, thus safeguarding user privacy. The initial phase involves conducting a thorough analysis of design requirements, drawing from existing literature [3, 4] to identify which privacy elements require protection. Subsequently, the project will implement computer vision models within the iOS app to identify target objects (e.g., human faces) captured by the camera. Finally, real-time occlusion will be applied to the pixels corresponding to detected target objects, thereby preserving privacy. Depending on progress, the project may incorporate additional features such as customizable options and accessibility enhancements, followed by remote mock testing of the app. The app created in this project will be used for further user studies, and the findings will be submitted to a prestigious computer science conference.

  By undertaking this project, students will accomplish several key learning objectives. Firstly, they will gain proficiency in developing iOS applications using the Swift programming language within the Xcode environment. Secondly, students will acquire skills in deploying machine learning APIs provided by Apple to integrate computer vision models into iOS apps. Lastly, they will learn how to leverage GPU capabilities to execute real-time rendering of desired colors on iOS devices. This project offers students an invaluable opportunity to deepen their understanding of mobile system concepts, tools, and cutting-edge computer vision technologies.

  References Cited

  1. https://aira.io/
  2. https://www.bemyeyes.com/
  3. Stangl, A., Shiroma, K., Davis, N., Xie, B., Fleischmann, K. R., Findlater, L., & Gurari, D. (2022). Privacy concerns for visual assistance technologies. ACM Transactions on Accessible Computing (TACCESS), 15(2), 1-43.
  4. Akter, T., Ahmed, T., Kapadia, A., & Swaminathan, S. M. (2020). Privacy considerations of the visually impaired with camera based assistive technologies: Misrepresentation, impropriety, and fairness. In Proceedings of the 22nd International ACM SIGACCESS Conference on Computers and Accessibility (pp. 1-14).

• Dr. Wei Zhang

  

  Faculty Profile

  View Dr. Wei Zhang’s faculty profile

  Research Areas:

  • Computer architecture
  • Compilers
  • Real-time systems

  Sample Project Description:

  Cybersecurity in Machine Learning-Based Internet of Things Devices (Drs. Wei Zhang and Soodeh Atefi)

  With the widespread use of Internet of Things (IoT) devices (e.g., smart homes or medical applications), the security of these devices has become increasingly important. Ensuring security is crucial for protecting individuals’ privacy and preventing unauthorized access to personal information. Since IoT devices play a crucial role in controlling and monitoring essential functions, compromised security could lead to malfunctions or unauthorized control, posing risks to health and safety [1]

  Modern IoT devices use machine learning (ML) to automate functionalities for users and leverage multiple ML-based data analysis techniques. However, this makes them vulnerable to adversarial attacks that target ML models where data is generated in IoT applications [2, 3].

  Adversarial attacks and their countermeasures have been widely studied in computer vision applications [4] (where the datasets are images) in contrast to IoT. Adversarial attacks can manifest during the training phase (data poisoning attacks) or during the model evaluation phase (evasion), presenting distinct challenges for detecting and filtering them. These attacks reduce the accuracy of the model (evasion attacks) or cause misclassification for specific examples (data poisoning attacks).

  In this project, an REU student will initially explore adversarial attacks and defenses against ML models trained on computer vision datasets to gain familiarity with the technical terms in the literature, as well as available attack and defense techniques, and assess the applicability of these techniques when trained on datasets generated in IoT applications (which differ from image data). Subsequently, the student will delve into adversarial attacks targeting ML-based IoT devices and implement ML-based defense approaches against these attacks. The experimental results will showcase the most effective defense approach capable of successfully detecting and filtering attacks, thereby securing IoT applications.

  .

  References Cited

  1. Malhotra P, Singh Y, Anand P, Bangotra DK, Singh PK, Hong WC. Internet of things: Evolution, concerns and security challenges. Sensors. 2021 Mar 5;21(5):1809.

  2. Singh A, Sikdar B. Adversarial attack and defence strategies for deep-learning-based IoT device classification techniques. IEEE Internet of Things Journal. 2021 Dec 27;9(4):2602-13.

  3. Albaseer A, Abdallah M, Al-Fuqaha A. Exploiting the Divergence Between Output of ML Models to Detect Adversarial Attacks in Streaming IoT Applications. InICC 2023-IEEE International Conference on Communications 2023 May 28 (pp. 3090-3095). IEEE.

  4. Bajaj A, Vishwakarma DK. A state-of-the-art review on adversarial machine learning in image classification. Multimedia Tools and Applications. 2023 Jun 17:1-66.

• Eligibility Information

  To be eligible for this REU program, you must:

  • U.S. citizen, U.S. national, or permanent resident of the United States.
  • At least 18 years old by the start of the program date.
  • Enrolled in an undergraduate program and majoring in Computer Science, Computer Engineering, or a closely related field.
  • Has completed at least one year of study (rising seniors preferred), and have at least one more semester to graduate after this program ends.
  • Minimum GPA of 3.0.
  • Will join the program in person to work 40 hours/week, cannot work at another job or enroll in class during the program.

  Women, underrepresented minorities, and students from institutions with limited research opportunities are especially encouraged to apply!

• Stipend, Housing, and Travel

  Stipend and Meal

  • Participants will receive $7,200 stipend and meal allowance for this 9-week program.

  Housing

  • Free on-campus housing (or housing allowance) will be provided.

  Travel

  • $492 travel allowance to and from the UofL campus will be provided. Make sure you keep your receipts!

  Payment

  • Students will receive their compensation in three equal payments: one at the beginning, one in the middle, and one at the end of the program.