EE Seminar: Fewer, Safer Bits: Improved Methods for Massive Storage
Electrical Engineering Systems Seminar
Speaker: Moshik Lanir Hershcovitch
Ph.D. student under the supervision of Prof. Adam Morrison and Prof. Boaz Patt-Shamir
Sunday, 30th March 2025, at 15:00
Room 011, Kitot Building, Faculty of Engineering
Fewer, Safer Bits: Improved Methods for Massive Storage
Abstract
Massive storage systems are the backbone of our digital infrastructure, enabling the seamless handling of enormous data sets across diverse applications. These systems face multiple, interrelated challenges in managing ever-growing volumes of data while balancing efficiency and security requirements. We address several such fundamental challenges through four key contributions:
1. In-Memory Database Efficiency: We introduce Elastic Indexes, a dynamic framework that automatically tunes the trade-off between storage space and query performance. This approach significantly reduces memory overhead in in-memory DBMS pipelines by optimizing indexed key storage while maintaining query performance through adaptive space-time trade-offs.
2. Lossless Compression for AI Models: We present ZipNN, an innovative compression framework that recognizes and exploits the unique characteristics of neural network parameters. This specialized compression technique significantly reduces storage requirements without compromising model accuracy or inference performance.
3. Key-Value Store Security: We identify novel key disclosure timing attacks in LSM-tree key-value stores utilizing range filters. Our prefix siphoning attack leverages benign queries for non-present keys to identify prefixes of actual keys—and in some cases, full keys—in scenarios where brute force searching for keys (via exhaustive enumeration or random guesses) is infeasible.
4. Relational Database Security: We extend our security analysis to database access controls, uncovering timing-based side channels in row-level security (RLS) implementations. This work demonstrates how query execution timing in multi-tenant environments can leak protected data.
השתתפות בסמינר תיתן קרדיט שמיעה = עפ"י רישום שם מלא + מספר ת.ז. בדף הנוכחות שיועבר באולם במהלך הסמינר
