[Paper Review] Reasoning over Public and Private Data in Retrieval-Based Systems

Simran Arora, Patrick A. Lewis+3 more|arXiv (Cornell University)|Mar 14, 2022

Privacy-Preserving Technologies in Data4 citations

TL;DR

This paper introduces the Public-Private Autoregressive Information Retrieval (PAIR) framework for privacy-preserving retrieval over both public and private data, proposes the ConcurrentQA benchmark for multi-distribution, multi-hop question answering, and demonstrates that existing systems face significant privacy-performance tradeoffs in this setting, which can be mitigated through selective prediction and concurrent retrieval over multiple data distributions.

ABSTRACT

Users and organizations are generating ever-increasing amounts of private data from a wide range of sources. Incorporating private data is important to personalize open-domain applications such as question-answering, fact-checking, and personal assistants. State-of-the-art systems for these tasks explicitly retrieve relevant information to a user question from a background corpus before producing an answer. While today's retrieval systems assume the corpus is fully accessible, users are often unable or unwilling to expose their private data to entities hosting public data. We first define the PUBLIC-PRIVATE AUTOREGRESSIVE INFORMATION RETRIEVAL (PAIR) privacy framework for the novel retrieval setting over multiple privacy scopes. We then argue that an adequate benchmark is missing to study PAIR since existing textual benchmarks require retrieving from a single data distribution. However, public and private data intuitively reflect different distributions, motivating us to create ConcurrentQA, the first textual QA benchmark to require concurrent retrieval over multiple data-distributions. Finally, we show that existing systems face large privacy vs. performance tradeoffs when applied to our proposed retrieval setting and investigate how to mitigate these tradeoffs.

Motivation & Objective

• To address the lack of privacy-aware retrieval systems that can simultaneously access public and private data for personalized applications.
• To identify that current multi-hop retrieval systems expose private data when retrieving from public sources, creating a critical privacy vulnerability.
• To propose a new benchmark, ConcurrentQA, to evaluate systems on concurrent retrieval over multiple data distributions (public and private).
• To investigate privacy-performance tradeoffs in retrieval-based systems when handling data across different privacy scopes.
• To explore methods such as selective prediction and concurrent retrieval to mitigate privacy leaks while maintaining performance.

Proposed method

• Proposes the PAIR (Public-Private Autoregressive Information Retrieval) framework to formalize retrieval over multiple privacy scopes, ensuring private data is not exposed during public retrieval.
• Designs ConcurrentQA, the first multi-hop question-answering benchmark requiring retrieval from both public and private data distributions, with questions that necessitate information from both sources.
• Employs iterative, beam-search-based multi-hop retrieval where each hop uses the question and prior retrieved documents to retrieve the next, but enforces privacy constraints on the sequence of retrievals.
• Introduces selective prediction as a mitigation strategy, where the system avoids retrieving public data if private data is involved in earlier hops, reducing exposure risk.
• Develops concurrent retrieval mechanisms that allow simultaneous or parallel access to public and private corpora, avoiding the need to expose private data to public systems.
• Uses a combination of sparse (BM25) and dense retrieval models to evaluate performance across different retrieval strategies in the new benchmark.

Experimental results

Research questions

• RQ1 How can retrieval-based systems be designed to reason over both public and private data without compromising user privacy?
• RQ2 What are the key privacy risks introduced by autoregressive multi-hop retrieval when private data is retrieved before public data?
• RQ3 How do existing retrieval models perform on a benchmark requiring concurrent access to public and private data distributions?
• RQ4 What tradeoffs exist between privacy preservation and retrieval performance in multi-source, multi-distribution settings?
• RQ5 Can selective prediction or concurrent retrieval strategies effectively reduce privacy leaks while maintaining high performance?

Key findings

• Existing multi-hop retrieval systems face significant privacy risks when retrieving private data before public data, as private documents may be exposed during subsequent public retrieval steps.
• The ConcurrentQA benchmark reveals that state-of-the-art retrieval models, including dense and sparse methods, perform poorly on concurrent public-private retrieval tasks, with performance drops of up to 75% compared to single-distribution settings.
• Sparse retrieval models like BM25 outperform dense models on certain ConcurrentQA questions, suggesting that semantic matching may not always be optimal in privacy-constrained settings.
• Selective prediction significantly reduces privacy leakage by avoiding public retrieval when private data is involved, though at the cost of reduced recall and performance.
• Concurrent retrieval over multiple distributions improves performance and reduces privacy risks compared to sequential retrieval, especially in multi-hop scenarios.
• The study demonstrates that current benchmarks are inadequate for evaluating public-private retrieval, necessitating new evaluation frameworks like ConcurrentQA to assess real-world privacy and performance tradeoffs.