Speaker: Paul Grubbs

Title: Breaking and Building End-to-End Encrypted Systems

Abstract: Today’s computer systems and their owners fail to protect data, and this failure is exacerbated by new threats caused by the explosion of cloud computing.  The consequences are dire: sensitive information like financial statements, medical records, and private messages are disclosed to malicious parties.  In my research at the intersection of security, cryptography, and systems, I work to change this by breaking and building efficient end-to-end (E2E) encrypted systems, which protect data by encrypting it throughout processing and storage. In this talk, I'll explain some of the flaws I've found in existing E2E-encrypted systems deployed to billions of users, and how the flaws have led me to a new methodology for building E2E-encrypted systems that's rooted in co-design of cryptography and systems. I'll conclude by outlining this methodology and some of the new E2E-encrypted systems I've built with it.

Speaker: Jack Hessel

Title: Unsupervised Discovery of Multimodal Links in Multi-image, Multi-sentence Documents

Abstract: Images and text co-occur constantly on the web, but explicit links between images and sentences (or other intra-document textual units) are often not present. We present algorithms that discover image-sentence relationships without relying on explicit multimodal annotation in training. We experiment on seven datasets of varying difficulty, ranging from documents consisting of groups of images captioned post hoc by crowdworkers to naturally-occurring user-generated multimodal documents. We find that a structured training objective based on identifying whether collections of images and sentences co-occur in documents can suffice to predict links between specific sentences and specific images within the same document at test time.
This is joint work with Lillian Lee and David Mimno

Speaker: Praveen Kumar

Title: Towards predictable network performance

Abstract: Performance isolation is a fundamental challenge in any shared system. While it has been well-studied in the context of operating systems, it is exacerbated in the context of shared public clouds as the scale and distributed nature of the cloud pose new challenges. In this talk, I will focus on the fundamental trade-off between network performance isolation and resource efficiency in public clouds, and demonstrate how isolation break down at the end-hosts can lead to unpredictable performance. Then, I will present a system, PicNIC, that navigates this trade-off to provide predictable performance by introducing the abstraction of a "predictable virtualized NIC" for each virtual machine (VM). PicNIC defines network performance objectives for each VM and leverages two key design principles to achieve them: (i) resource sharing based on performance objectives and (ii) applying backpressure to sources. Finally, I will conclude with some thoughts on how we can achieve both high and predictable network performance as we move towards domain-specific hardware and accelerators for networking.