Deepak Gouda

Resource Public Key Infrastructure (RPKI) is a critical component in securing the inter-domain routing infrastructure today. More than 50% of the routed IPv4 and IPv6 prefixes are covered by RPKI Route Origin Authorizations (ROAs). ROAs are cryptographically verifi- able records of the Autonomous System (AS) authorized to originate routes to a set of prefixes. Network operators are increasingly rely- ing on RPKI to validate routing information and reduce the spread of BGP hijacks and misconfigurations. RPKI infrastructure has five root authorities maintained by the five Regional Internet Registries (RIRs). Each root authority independently implements its RPKI in- frastructure, choosing how to manage certificate production from its self-signed root of trust certificate. In this poster, we study the different designs of RPKI infrastructure across the five roots and how these differences impact the characteristics of the RPKI Cer- tificate repository, such as scalability and compute requirements. We discover that some RPKI repositories are computationally more expensive than others due to their design.

We present a new on-device pipeline that efficiently summarizes lecture videos and provides relevant answers directly from a smartphone. We utilize widely accessible tools like OCR and Vosk speech-to-text, coupled with powerful large language models (LLMs), to identify crucial sentences and generate summaries. By harnessing the capabilities of LLMs and the computational power of mobile devices, we fine-tune and quantize BERT and GPT-2 to achieve efficient lecture video summarization and question answering on consumer-grade smartphones like the Pixel 8 Pro. Notably, this approach eliminates the need for cloud APIs, ensuring enhanced user privacy and minimal mobile data usage.https://www.youtube.com/shorts/zwGdONlKays

The optimal learning rate for adaptive gradient methods applied to λ-strongly convex functions relies on the parameters λ and learning rate η. In this paper, we adapt a universal algorithm along the lines of Metagrad, to get rid of this dependence on λ and η. The main idea is to concurrently run multiple experts and combine their predictions to a master algorithm. This master enjoys O(d log T) regret bounds.