UROC

"The Midas Touch" - Human-Computer Interaction

Project Leader: Professor Dan Vogel

The hope for camera-based input devices like the Microsoft Kinect and LEAP Motion is that interacting with computers becomes more natural. But in practice, waving your arm and pointing your finger to navigate something like Netflix can leave a lot to be desired. In this mini-project, we’ll work on a fundamental Human-Computer Interaction problem associated with all computer vision-based input, the “Midas Touch Problem.” Like the Greek Myth where everything King Midas touches turns to gold, with computer vision, everything you do may be interpreted as input; even if you’re just waving at a friend. The challenge is to find techniques that work well with noisy computer-vision tracking, aren’t too tiring, and are easy and fast to perform. We’ll review current approaches and go through a mini bootcamp for computer vision coding so we can prototype and test new interaction techniques — and experience first hand what Human-Computer Interaction research is like.

"The Internet of Things"

Project Leader: Professor Srinivasan Keshav

With the rapid decline in the costs of computers and sensors, it is possible to build small-scale sensor systems that sense and report on the physical world. Examples are sensors that measure room temperature, lighting level, and air quality. Related to these sensors are computer-controlled actuators, that effect changes in the physical world by, for example, controlling lights or appliances. These are the foundation of the ‘Internet of Things’ (IoT). In this mini-project, you will learn how to work with the Onion Omega IoT device to measure room occupancy and light level and use these to intelligently control a light bulb. This will allow you to learn about sensing, communication, computing, and actuation; skills that you can bring to other projects that may interest you in the future

"Adversarial Training and Security of Modern Machine Learning Algorithms" - Machine Learning

Project Leaders: Professor Yaoliang Yu and Professor Pascal Poupart

Machine learning is the science of extracting knowledge from massive amount of potentially high-dimensional data. In recent years we have witnessed enormous successes in applying machine learning algorithms to attack challenging real-life problems, such as image and speech recognition, machine translation, autonomous driving, personalized medicine, recommendation systems, computer games, etc. All of this would not be possible without data, the "oblivious" input to machine learning algorithms. In this mini-project, we will examine to which extent data can affect a particular family of learning algorithms (such as deep neural networks) in two seemingly contradictory senses: (a) Can we use generative adversarial networks to generate additional data similar to our training data? Would incorporating such additional data improve the performance of a learning algorithm? (b) Can we break down a learning algorithm by feeding into it adversarially generated data? If yes, how can we effectively generate such adversarial data? Can a learning algorithm detect such adversarial data?

"Analyzing Tweets with Spark" - Data Mining

Project Leaders: Professor Jimmy Lin and Professor Ihab Ilyas

Over the past few years, we have seen the emergence of "big data": disruptive technologies that have transformed commerce, science, and many aspects of society. These developments are enabled by infrastructure that allows us to distribute computations across hundreds or even thousands of commodity servers. One key breakthrough that makes this all possible is the development of abstractions for data-intensive computing that allow programmers to reason about computations at a massive scale, hiding low-level details such as synchronization, data movement, and fault tolerance. Spark is one such data processing platform. In this mini-project, we'll explore what a "data scientist" does today in analyzing big data. For concreteness, we'll focus on tweets.

"Attacking Searchable Encryption" - Cryptography, Security and Privacy (CrySP)

Almost everybody is storing data on the cloud, be it using a mobile phone or using web-based email. Wouldn't it be nice to store your data in the cloud, but it be encrypted such that no one including the cloud provider can read it? What sounds simple at first prevents cloud applications from performing their work, such as searching through the document collection. Therefore researchers have been working on encryption schemes that are also searchable. However, their security is unclear even to the scientific community. In this project we will analyze the security of searchable encryption schemes. We will implement attacks on the cryptographic schemes showing how easy (or difficult) it is to break them. We start by recreating known attacks in order to understand the weaknesses of the encryption schemes. Then we will continually refine our approach and our goal is to develop even more powerful attacks.