Zahra Ghodsi
UCSD
zghodsi@ucsd.edu
Bio
Zahra is a postdoctoral scholar in the ACES Lab at University of California San Diego. She received her PhD from New York University under the supervision of Prof Siddharth Garg in January 2021. Her research lies at the intersection of security privacy and machine learning and she maintains research interest in applied cryptography. She was awarded the NYU Inclusive Excellence Award the J.P. Morgan AI Fellowship and the NYU Ernst Weber Fellowship during her doctoral studies. She completed an internship at NVIDIA Research and has served on the program committee of NDSS 2022 and artifact evaluation committee of ASPLOS 2021.
Secure Frameworks for Outsourced Deep Learning
Secure Frameworks for Outsourced Deep Learning
Deploying modern machine learning models in real world applications comes with several challenges. First large amounts of data is required for training which can be difficult to obtain and second training and performing inference can be computationally expensive. Collaborative computing and machine learning as a service (MLaaS) solutions allow different parties to benefit from these applications but they also raise immediate security concerns relating to the integrity (or correctness) of computations and the privacy of parties’ assets. My research aims to build efficient frameworks to provide rigorous privacy and integrity guarantees for collaborative machine learning. I have built frameworks for secure outsourced computations of deep learning inference that rely on techniques in verifiable computing and secure multi-party computation from cryptography. The challenge in incorporating cryptographic protocols is that they typically incur large overheads which coupled with the size and scale of deep learning computations renders them impractical. The key idea behind the frameworks I developed is co-designing cryptographic protocols and machine learning models to bring these systems closer to practicality.