Private AI

We present recent results of the work performed for the DARPA Cooperative Secure Learning (CSL) program. The main goal of the CSL program is to develop methods for secure machine learning in scenarios with multiple data owners. This talk focuses on large-precision comparison and arbitrary function evaluation using FHEW/TFHE bootstrapping. A comparison of two encrypted numbers is an important operation needed in many machine learning applications, for example, decision tree or neural network inference/training. An efficient instantiation of this operation in the context of fully homomorphic encryption (FHE) can be challenging, especially when a relatively high precision is sought. We propose and implement in software algorithms the computational complexity of which scales logarithmically with ciphertext (plaintext) modulus, enabling the support of large-precision comparison and arbitrary function evaluation in machine learning applications.

<p>This is joint work with Craig Gentry and Vadim Lyubashevsky</p><p>Non-interactive publicly verifiable secret sharing (PVSS) schemes allow
parties to re-share a secret in a decentralized setting in the presence
of malicious parties. A recently proposed application of PVSS schemes is
to enable permissionless blockchains to “keep a secret” via a sequence
of committees that share that secret. Such committees can use the secret
to produce signatures on the blockchain’s behalf, or to disclose hidden
data conditioned on consensus that some event has occurred. Such a
setting may involve thousands of parties, so the PVSS scheme that it
uses must be very efficient, both in computation and communication. Yet,
previous PVSS schemes have large proofs and/or require many
exponentiations over large groups.</p><p>

We present a non-interactive PVSS scheme in which the underlying
encryption scheme is based on the learning with errors (LWE) problem.
While lattice-based encryption schemes are very fast, they have issues
with bandwidth (long ciphertexts and public keys). We deal with the
bandwidth issue in two ways.</p><p>

* First, we adapt the Peikert-Vaikuntanathan-Waters (PVW) encryption
scheme to the multi-receiver setting so that the bulk of the parties’
keys is a common random string. This allows us to get a constant
amortized communication rate: The plaintext/ciphertext expansion is
about 1/60 for 100 parties, and about 1/8 for 1000 parties, approaching
1/2 as the number of parties grows.</p><p>

* Second, we use bulletproofs over a DL-group of order about 256 bits to
get compact proofs of correct encryption of shares. Switching from
the lattice setting to the DL setting is relatively painless, as we
equate the LWE modulus with the order of the group, and apply dimension
reduction to vectors before the switch to minimize the number of
exponentiations in the bulletproof.</p><p>

An implementation of our PVSS for 1000 parties showed that it’s
quite practical, and could remain so with up to a two order of magnitude
increase in the group size.</p>

The K-Statistics System refers to a Korea’s national statistical system for data linkage and data security. Amendment of Korea’s Statistics Act is currently underway to establish the legal basis to link statistical registers (hub) and administrative data (spokes) that are scattered across government agencies. To guarantee security of linked data, Statistics Korea so far primarily depends on hardware protection by allowing data linkage only in physically protected data centers. Afterwards, Statistics Korea will leverage homomorphic encryption, differential privacy, and synthetic data to ensure security of such linked data even outside of the data centers. A pilot project has been implemented to perform a logistic regression analysis of the survival of small businesses by linking data in their homomorphically encrypted state between Statistics Korea and Gyeonggi Province. The pilot project demonstrates that homomorphic encryption and key sharing has potential to facilitate data cooperation between different government agencies who don’t necessarily trust each other.

Key Words: K-Statistics System, Homomorphic Encryption, Differential Privacy, Synthetic Data, Survival Analyses of Small Businesses

This talk first reviews state-of-the-arts in many NLP tasks, that rely on the language understanding capability of increasing larger pre-trained language models. We then discuss privatization mechanism for embeddings based on homomorphic encryption, to prevent potential leakage of any piece of information in the process of text classification. Our method enables the training of a strong classifier within hours. 

In this talk will present the Exposure Notifications Private Analytics (ENPA) system developed jointly by Google and Apple to provide health authorities with privacy preserving metrics in the Exposure Notifications System for COVID-19. We will look at the problem of privacy preserving measurement more broadly and compare and contrast existing techniques.