Doctoral Speaking Skills Talk - Siddharth Jayashankar
May 4, 2026 11:30AM—12:30PM
Location:
6501
-
Gates and Hillman Centers
Speaker:
SIDDHARTH JAYASHANKAR,
Ph.D. Student, Computer Science Department, Carnegie Mellon University
https://sidjay10.github.io/index.html
Fully Homomorphic Encryption (FHE) is a transformative cryptographic primitive that enables direct computation on encrypted data, offering robust security guarantees for sensitive workloads. By ensuring user data remains encrypted throughout the entire execution pipeline, FHE facilitates secure inference even in the presence of untrusted model providers or malicious attackers. However, despite this potential, practical adoption has been hindered by massive computational overheads. While ASICs have been proposed to mitigate these costs, they rely on advanced semiconductor manufacturing and high-end packaging, rendering them currently impractical for widespread production.
In this talk, I will present my work on Cardamom, an FHE framework that achieves ASIC-competitive performance using commodity datacenter GPUs. Cardamom introduces a holistic approach to optimization across the compute, memory, and communication stacks to automatically generate high-performance implementations for FHE programs—ranging from small-scale CNNs to Large Language Models (LLMs) like Llama3-8B. Our evaluation demonstrates that Cardamom outperforms expert-optimized libraries by up to 2.25 times and achieves a milestone in FHE performance: the first sub-10ms CKKS bootstrapping at 7.5ms. Furthermore, Cardamom enables end-to-end encrypted inference for BERT-Base in 8 seconds and Llama3-8B in 43 seconds, representing a significant leap toward the practical deployment of secure, large-scale AI.
Presented in Partial Fulfillment of the CSD Speaking Skills Requirement
Contact
Matt Stewart