2020 Intern Presentations
ZOOM INFO Available upon request (password required) - Email [email protected]
Thursday, Aug. 20, 1 p.m. Central Time
Presenters: Kuheli Sai, Namratha Urs and Sam Usman
Kuheli Sai, returning GHC 2019 intern
Mentors: Qian Gong/Mine Altunay (OCIO)
Research Project: Few-Shot Learning for Network Traffic Anomaly Detection
Abstract: Network Intrusion Detection System (NIDS) often serves as the first line of defense against cybercrime. Due to the prevalence of traffic encryption and frequent changes of malware patterns, network researchers and engineers have recently proposed to combine behavior modeling with machine learning (ML) techniques for intrusion detection. Machine learning methods typically require large number of training samples to avoid overfitting. However, the recorded attack data is rare, and new types of malicious activities come out daily. We therefore formulate the network traffic classification as a problem of few/one shot learning, where a neural network is trained to extract information from small quantities of training samples, and quickly adapt to the changes in non-stationary input distribution.
Namratha Urs: GHC 2020 intern
Mentor: Marco Mambelli, SCD
Research Project: On-Demand Provisioning of CernVM File System with GlideinWMS
Abstract: High Energy Physics entails an abundance of computing resources. This requirement is fulfilled by local batch farms, grid sites, private/commercial clouds and supercomputing centers. Particle physics collaborations use the CernVM File System as a scalable, reliable and low-maintenance software distribution service. The CernVM File System (CernVM-FS or CVMFS) is a write-once, read-everywhere distributed file system based on HTTP. CVMFS is used to deploy scientific software to thousands of nodes — virtual machines and physical worker nodes — on a worldwide distributed computing infrastructure. Experiments use CVMFS for: (1) distributing experiment software and data (calibrations and simulations), and (2) facilitating containerization for regular (unprivileged) users by hosting conteiners images efficiently. Most sites provide a local installation of CVMFS, however, some sites (especially, HPC resources) do not.
GlideinWMS is a workflow manager that facilitates provisioning of different kinds of resources for distributed computing. This is done by acquiring the resources and sending the pilot jobs (glideins) that test and setup all worker nodes.
This work extends GlideinWMS; specifically it adds to the glidein the ability to provide CVMFS on sites where it is not locally available. By evaluating the site, this new glidein determines the most reliable option to provision CVMFS on the worker node. The added availability of CVMFS also allows user jobs to be started inside Singularity containers even when Singularity is not installed on the site. The result is a lower overhead for site administrators, that have less software to install, and the ability for user jobs to run [almost all the times] in the desired container image and access the software and data distributed via CVMFS, making life easier for the scientists. This is essential to use GlideinWMS on HPC resources.
Intern: Sam Usman: GHC 2020 intern
Mentor: Brian Nord, SCD
Research Project:Simulating Light From the Early Universe
Abstract: The big bang started the universe, and the dense, hot universe cooled as it expanded. At around 370,000 years, it cooled enough to be transparent to optical light, which then stretched into microwaves as the universe continued to expand. Analyzing the cosmic microwave background (CMB) can and has revealed key information about the early universe, such as the distribution of dark matter, dark energy and baryonic matter. We can still learn about the initial expansion of the universe if we can identify a polarization pattern called B modes, which would indicate gravitational waves had been created during the inflationary period of the universe. Unfortunately, this CMB is distorted on the scale of B-modes by gravitational lensing. To efficiently disentangle these signals, we turn to machine learning algorithms and other statistical methods, like simulation-based inference. To train these complex algorithms, I spent my summer creating a set of tools to create simulations of the CMB with gravitational lensing.