I'm currently a M.S. student in the Computer Science Department at Johns Hopkins University. I'm a member of the Distributed Systems and Networks lab, working with prof. Yair Amir. My main interests are in networks and distributed systems, but I also enjoy computer security, graphics, and parallel architectures. I received my undergraduate degree in Computer Engineering at Johns Hopkins in 2010. Afterwards, I decided to pursue a M.S. in Computer Science.

Clouds offer a cost effective approach for many applications, including storage, messaging and computing resources. As more of the world shifts their applications from privately managed entities to the cloud service model, the global IT infrastructure will become increasingly dependent on a small number of very large distributed systems. As a result, the number of targets for malicious attacks drastically decreases. Therefore, the managers of these clouds are responsible for providing impeccable reliability, availability and security of their widely used infrastructure. While the reliability and availability issues have been tackled in the past, cloud security is still an underdeveloped topic. If parts of today’s clouds are compromised, the entire system can be brought down, completely ruining all applications and services running on top of the network. Thus, ensuring the resiliency and performance metrics of cloud systems is of utmost importance.

Traditionally, security measures aim to provide two main goals: prevention and detection. We argue that this conventional approach is insufficient. Modern cloud systems entail a large concentration of homogeneous hosts on the widely accessible Internet. With essentially the same software running on each of these hosts, the capability to compromise one cloud node usually translates to the capacity to take over the entire system. The implicit trust among hosts and lack of strong internal checks make matters worse. With this in mind, completely preventing clouds systems from being compromised seems impossible. Therefore, we believe that intrusion tolerance, the ability of a system to maintain functionality even when some of its members have been compromised, is the most valuable security guarantee. We have developed four major areas of focus that we feel will provide an adequate solution to the cloud resiliency problem: scalable intrusion-tolerant replication, tunable intrusion-tolerant messaging, diverse attack surfaces and real-time detection and prediction. We are interested in the design, architecture, implementation and deployment of a real cloud networking system that contains the aforementioned resiliency traits, ultimately providing the first intrusion-tolerant cloud of its kind.

Currently, I am working with a team of students and professors from Johns Hopkins University, Purdue University and the University of Virginia. I have put the majority of my effort into two of the focus areas mentioned earlier, namely the intrusion-tolerant replication and intrusion-tolerant messaging. For the replication, we aim to design and implement both a scalable and performance guaranteeing replication protocol in the presence of malicious nodes. For the messaging, we have designed several routing protocols that can overcome Byzantine neighbors, each with its own guarantees and tradeoffs. We are currently in the analysis and implementation phase of these protocols. (Funding: MRC)

Most of my previous research has been in one of two domains: Distributed Systems and Low-Level Android Development. I have also done some research involved with computer vision and surgical robotics.

Spiny Android

The goal for this project was to develop software to run on Android phones that would share connectivity (WIFI, BlueTooth, etc) between nearby phones running the same software. In the end, we ported Spines directly onto the Android phone's hard disk (no virtualized application). The resulting proof-of-concept resembled something of a mesh network. Project page.

Remote Telesurgery

This project aimed to develop extremely low-latency and reliable communication for use in telerobotic remote surgery. In the target system, an expert robotic surgeon would perform surgery over long geographical distances, requring a low-latent and reliable two-way communication channel for robot commands and high-definition stereoscopic video. Overlay networks were used in order to cope with the everyday problems experienced by internet routing and provide the necessary latency and availability constraints. In addition, the da Vinci was used as the surgical system. Project page.

Automated Assessment for Robotic Surgery Training

In general, computing objective and quantitative measurements for surgical training is hard. Classical measurements include completion time, visual error, etc. But these measurements do not always accurately determine when a surgeon in training can graduate to the next level. Robotic surgery provides good ways of measuring in a level of detail that makes new training measurements possible. Using actual training video and kinematic data from the da Vinci surgical system, computer vision techniques were used to automatically annotate and segment trainee sessions. Both manual and automatic measurements could then be performed on the resulting segments.

2010 - John Boswell Whitehead Award for outstanding achievements by an undergraduate - Computer Engineering

Database Systems, Declarative Methods, Parallel Programming, Distributed Systems, Algorithms, Randomized Algorithms, Advanced Distributed Systems and Networks, Computer Graphics, Compilers and Interpreters, Security and Privacy in Computing

Intermediate Programming (Fall 2011, Spring 2012)

2010-2012 - Computer Science Happy Hour Czar - responsible for organizing weekly get togethers for staff, faculty, and students.