Modern cryptography, including multi-party computation, zero-knowledge
proofs and fully homomorphic encryption, offers provable privacy for
computations over sensitive data and enables a new generation of
applications. Applications span from privacy-preserving machine
learning and secure financial analytics to privacy-respecting
decentralized systems. Despite decades of progress, a persistent gap
remains between the theoretical efficiency of these protocols and the
performance demands of real-world deployments. This gap arises from a
disconnect between hardware-agnostic protocol design and how modern
hardware actually operates. Asymptotically optimal protocols routinely
rely on primitives that hardware executes slowly, and even hardware
aware protocols perform poorly when implementations ignore instruction
pipelines, memory hierarchy, and vector units. My work is motivated by
a core goal: to build secure computation primitives and systems that
are provably secure while remaining concretely practical on real
hardware.
In this talk, I will summarize my prior work spanning pseudorandom
correlation generators, garbled circuits, RAM-based MPC, oblivious
RAM, threshold FHE, and anonymous messaging. I will focus on three
examples of the co-design philosophy at the primitive level: a
maliciously secure distributed point function that closes the semi
honest–active gap at 50× less communication; a mixed-mode oblivious
RAM that allows public accesses and shaves the log factor and concrete
overheads for these operations, while matching the fully-private lower
bound; and a concretely efficient succinct garbling scheme that turns
rate-one garbled circuits from a theoretical curiosity into a
deployable tool. I will then present the forward-looking direction for
my ongoing and future research: extending the co-design methodology
from the CPU down to specialized accelerators, targeting large-scale
FHE workloads.
Jensen Smith
Email