Efficient Fuzzy PSI under One-Sided Assumptions

arXiv:2608.17770 · cs.CR · Submitted 2026-08-18 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Efficient Fuzzy PSI under One-Sided Assumptions".

Jane: The paper was written by Xinpeng Yang, Meng Hao, Yanxue Jia, Chenkai Weng, Yonggang Wen et al. from Nanyang Technological University, Singapore and Singapore Management University, Singapore and Illinois Institute of Technology, Chicago, Illinois, USA and Arizona State University, Tempe, Arizona, USA.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Summary: Tom: So we spent a good bit of time unpacking the title, which really focused on the efficiency gains and handling fuzzy data. Now, let's dig into the paper's summary—the core mechanism they’ve developed for "Efficient Fuzzy PSI under One-Sided Assumptions."

Jane: If I can simplify what they summarized, it seems like they found a way to build an encrypted bridge between two parties, letting them compare fuzzy data points without either side ever seeing the other's actual inputs.

Lu: The summary really hammered home that this isn't just about matching keys; it's about establishing proximity in a private space. They are creating a mathematical environment where 'close enough' is secure enough for computation.

Meng: My take from the summary is that the key breakthrough seems to be optimizing the resource requirements—the computational steps needed—to make this process fast enough for real-time use cases, which is often the biggest hurdle in cryptography.

Lalam: The implication of this mechanism is really about democratizing access to data insights. Currently, only massive entities with perfect data can run these types of analyses; this method opens it up to smaller organizations needing similar power.

Tom: That's right, Lalam, because the summary emphasizes how much better the performance is compared to older fuzzy PSI methods. It's not just that it works; it works *fast*.

Jane: Which means we can stop treating privacy as a slowdown and start treating it as an integrated feature—a natural part of computation. That’s a huge shift in thinking for industry adoption.

Lu: And considering the structure of their approach, I think this opens up possibilities for cross-domain research, like combining medical records from different regional clinics that use varying data standards.

Meng: So, if we're talking about practical deployment, are the system inputs standardized? I'd need to know if they recommend a specific data format or infrastructure setup for integrating this fuzzy matching capability into existing cloud services.

Lalam: Ultimately, what this paper gives us is the blueprint for a truly collaborative digital economy—one where value is derived from shared insights without sacrificing individual sovereignty over personal information.

Improvements: Tom: We’ve talked about the concept and the summary, and now we're looking at what improvements the authors suggest in "Efficient Fuzzy PSI under One-Sided Assumptions." It seems they are continually refining this already complex system.

Jane: When I read about these suggested improvements, it feels like they aren't just fixing bugs; they're making the entire architecture leaner and more robust against real-world imperfections we hadn’t accounted for before.

Lu: The incremental improvements they detail—especially regarding parameterization—suggest a much higher degree of flexibility. It means the system can be tuned for vastly different levels of data sparsity or noise, which is incredibly powerful.

Meng: Speaking of tuning, I noticed they discuss optimizing the communication complexity alongside the computation complexity. That dual optimization is crucial because in any distributed system, sending data back and forth often becomes the biggest bottleneck.

Lalam: What this means for society is that these improvements allow us to handle increasingly complex and heterogeneous data streams—think mixing genomic data with behavioral patterns—all while maintaining absolute privacy integrity.

Tom: Right, Jane, it's about taking something conceptually brilliant and making it practically robust enough to handle the chaos of real life. The focus

Paper discussion segment 3: Tom: So, building on our discussion about Fuzzy PSI and one-sided assumptions, what really exciting changes did these authors propose that make this whole framework better?

Jane: Basically, while we covered that fuzzy matching handles uncertainty well, these improvements really tighten up the mathematical guarantees around how much information can actually leak.

Lu: Exactly! The key breakthrough isn't just making it fuzzy; it’s making the *efficiency* of the guarantee better, which means we can process way more complex or larger datasets without sacrificing privacy robustness.

Meng: That sounds great on paper, Lu, but when you say "efficiency," are we talking about computational efficiency—meaning faster runtime—or resource efficiency, like needing less memory?

Tom: Hey, Meng raises a good point; the engineering side of this has to be solid! Jane, can you clarify how this improved efficiency impacts real-world deployment?

Jane: Think of it like this: if the old method was a huge, clunky machine that needed tons of power just to run the privacy checks, these improvements are tuning it up so it runs faster and needs far less overhead.

Lu: And because they've managed to improve the underlying assumptions related to one-sided knowledge, we can apply this technique in more niche areas where we only know certain types of data about a user, which is huge for personalized AI.

Meng: If I understand correctly, this means we don't need perfect data from both sides—just knowing what one side *definitely* knows—is enough to build secure systems?

Tom: It sounds like they’ve made the privacy mechanism more adaptable to imperfect information sources, which is exactly where most real-world data lives!

Jane: That's right, Tom. It moves PSI from a theoretical academic exercise into something genuinely practical for corporate use cases today.

Lu: This really opens the door for collaborative research across industries that previously couldn't share raw data because of those asymmetrical knowledge gaps.

Meng: So, if we could deploy this at scale, could it handle, say, federated learning scenarios where multiple hospitals contribute patient data but can't pool it all?

Lalam: The ability to build robust systems using incomplete or partially known information is fundamentally changing how humanity shares knowledge; it allows for unprecedented levels of collaboration without compromising individual autonomy.

Tom: That’s a huge leap forward, Lalam—it means we can finally get truly powerful AI models trained on diverse, sensitive global datasets!

Jane: It really solidifies the idea that privacy doesn't have to mean sacrificing utility anymore.

Lu: Speaking of utility, I wonder how this PSI framework integrates with differential privacy metrics?

Meng: Well, if we nail down the implementation details for data sharing, what’s the next major technical hurdle we need to tackle?

Conclusion: Tom: So, taking all that discussion about fuzzy PSI and one-sided assumptions, what really stings is how much this work advances the field of privacy-preserving AI overall.

Jane: Exactly, Tom; it shows us a really robust way to combine multiple advanced concepts—privacy, fuzziness, and these specific assumptions—into something actually usable in practice.

Lu: I think the implication here goes far beyond just data sharing; this framework could fundamentally change how organizations design their entire data infrastructure.

Meng: Right? Because if you can build this level of privacy protection into the core architecture, it radically lowers the regulatory risk for large-scale deployments that are currently stuck in legal review.

Lalam: And on a cultural level, giving businesses this much reliable privacy control fosters a deeper trust relationship with the users, which is absolutely crucial for the future adoption of advanced AI systems.

Tom: That’s a huge point, Lalam; it really shifts the focus from "what data can we take?" to "how much trust can we build?"

Jane: It makes privacy a feature you sell alongside functionality, rather than just an afterthought that costs extra time and money.

Lu: Honestly, I’m excited about how many new applications this opens up in fields like healthcare or personalized education where data sensitivity is paramount.

Meng: From an engineering standpoint, the biggest win is the *efficiency* they achieved; it means we don't have to sacrifice performance just to meet strict privacy requirements anymore.

Lalam: Ultimately, advancements like "Efficient Fuzzy PSI under One-Sided Assumptions" mean that technological progress and ethical responsibility can finally move forward together.

Tom: It’s a massive win for the entire AI community, genuinely changing the conversation around data utility.

Jane: Well, we gotta wrap up our deep dive on this fascinating paper today. We loved talking through these implications with all of you!

Lu: Thank you so much for having us; it was a blast exploring the possibilities of this research.

Meng: Thanks for letting us break down the practical side of things; I learned a ton about the real-world impact.

Lalam: It’s been an incredibly insightful discussion, and we appreciate you giving us a platform to discuss how these advances improve culture.

Tom: We really appreciate it! And that wraps up our look at "Efficient Fuzzy PSI under One-Sided Assumptions." Stay tuned because next week, we're turning our attention to something completely different...

Nanyang Technological University, Singapore · Singapore Management University, Singapore · Illinois Institute of Technology, Chicago, Illinois, USA · Arizona State University, Tempe, Arizona, USA

cs.CR

Submitted: 2026-08-18

Updated: 2026-09-13

Comments: Accepted to ACM CCS 2026

Code: https://github.com/Th0masAndy/FPSI-One-Sided

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 94/100

The gist: Fuzzy private set intersection (PSI) allows two parties to identify approximately matching elements between their input sets, where a match occurs if their distance is at most a threshold delta.

Key concepts

Fuzzy PSI: This mechanism enables two parties to find common elements in their datasets while maintaining privacy. It creates a mathematical environment where 'close enough' is secure enough for computation, allowing comparison of fuzzy or uncertain data points without revealing the actual inputs.
One-Sided Assumptions: This refers to the system's ability to function securely even when one party does not have perfect data. The framework is designed to be highly adaptable and robust, making it useful for real-world scenarios where information sources are incomplete or asymmetrical.
Efficiency Gains: The core improvements focus on optimizing both computation and communication complexity. This dual optimization ensures the process runs fast enough for real-time use, allowing privacy checks to be integrated into systems without sacrificing performance.

Terminology

Aggregation — 10× in this paper · explained in 1 episode
Core — 8× in this paper · explained in 2 episodes
Benchmarking — 2× in this paper · explained in 3 episodes
Benchmark — 1× in this paper · explained in 2 episodes
Communication Overhead — 1× in this paper · explained in 2 episodes
Computational overhead — 1× in this paper · explained in 2 episodes
Differential Privacy — 1× in this paper · explained in 4 episodes
Federated Learning — 1× in this paper · explained in 15 episodes
Generalization — 1× in this paper · explained in 13 episodes
Generative AI — 1× in this paper · explained in 2 episodes
Information Leakage — 1× in this paper · explained in 2 episodes
Robustness — 1× in this paper · explained in 5 episodes
Sparsity — 1× in this paper · explained in 4 episodes

Summary

Fuzzy private set intersection (PSI) allows two parties to identify approximately matching elements between their input sets, where a match occurs if their distance is at most a threshold delta. While significant progress has been made in this field, existing constructions for general Minkowski distances have historically relied either on strong two-sided geometric separation assumptions or they incur substantial overhead under one-sided assumptions. This work addresses these limitations by presenting the first concretely efficient fuzzy PSI protocols for general L p in [1, infinity] distances under one-sided assumptions, relying solely on lightweight symmetric-key primitives.

Core Methodology and Assumptions

The proposed protocols operate under weaker one-sided assumptions, meaning only one party’s inputs are required to satisfy certain constraints while the other may hold arbitrary inputs. The constructions support both sender-sided and receiver-sided settings. Key features of the methodology include:

Utilizing spatial hashing techniques that partition the space U d into cells of side length 2 delta.
Employing a custom oblivious programmable pseudorandom function (OPPRF) to facilitate matching without revealing information about non-matching elements.
Implementing a novel building block, the conditional selection functionality (FConSel), which allows for the secure and private selection of matches based on an equality or comparison check.

Logarithmic Complexity via Prefix Trie Techniques

A major limitation in prior works was their superlinear complexity in delta. To overcome this, the authors non-trivially incorporate prefix trie techniques into their protocols. This approach allows for a dimension-by-dimension filtering mechanism that prune[s] the exponential search space. As a result, the optimized protocol achieves logarithmic complexity O(delta) in the distance threshold delta for general L p distances, significantly improving upon previous complexities of O((delta)d) or O(delta).

Support for General Distances and Sparser Distributions

The protocols are designed to handle the full range of Minkowski distances, supporting general L p in [1, infinity] metrics. Furthermore, the authors extend their scope to address sparser input distributions. In this specialized setting, they propose more efficient protocols tailored to the case where inputs are less densely distributed in space. This is achieved by reversing the roles of the sender and receiver in the spatial hashing procedure, which balances workloads and yields improved concrete performance.

Performance and Practical Impact

Extensive experiments demonstrate that these protocols significantly outperform prior works under comparable assumptions. The improvements are substantial across various parameter settings:

Against van Baarsen and Pu (EUROCRYPT’24), the protocols achieve up to 239 times faster computation and up to 20 times lower communication.
Against Dang et al. (CCS’25), they achieve up to 518 times speedup and up to 63 times communication reduction.
Against Bui et al. (ASIACRYPT’25), the performance is even more dramatic, showing up to 4818 times faster computation and up to 282 times lower communication.

The overall efficiency of these protocols stems from their reliance on symmetric-key primitives, avoiding both the superlinear complexity of prior works and the expensive public-key operations associated with other approaches.

Improvements for AI systems

Improvements to AI Systems:

Implementation of Logarithmic Complexity Matching Modules (O(delta)): We replace standard quadratic or linear search mechanisms for approximate set intersection with modules leveraging prefix trie techniques (FConSel). This dramatically reduces the computational overhead associated with handling high-precision distance thresholds (delta).
General Metric Support (L p Distance Handling): Our systems can support fuzzy matching using any L p norm (p in [1, infinity]), moving beyond the limitations of L infinity. This allows the AI to process diverse and naturally noisy data distributions (e.g., Manhattan distance for certain sensor data).
Flexible Privacy Constraints (One-Sided Assumptions): The core architecture supports one-sided assumptions (sender-sided or receiver-sided). This enables deployment in real-world scenarios where one party's input distribution is arbitrary, without requiring the costly and often unrealistic two-sided geometric separation constraints.
Symmetric Primitive Integration: We utilize lightweight symmetric-key primitives (OPPRF, EqRand) instead of heavy public-key infrastructure (AHE). This integration significantly lowers the computational and communication footprint compared to prior state-of-the-art solutions.
Sparse Data Optimization: Specific protocols are tailored for sparser input distributions, enabling highly efficient processing when the input feature sets are not uniformly dense.

What the Improved AI System Can Do:

Perform Privacy-Preserving Biometric/Semantic Matching at Scale: The system can identify approximate matches between two parties' data (e.g., comparing noisy biometric templates or semantic embeddings) without revealing any information about elements that do not match, even when the data is distributed arbitrarily.
Process High-Precision Feature Vectors Rapidly: Due to the O(delta) complexity, it can perform fuzzy matching on high-dimensional feature vectors (d-dimensional space) with very fine distance thresholds (delta) at a speed that scales logarithmically with the threshold, ensuring real-time performance even when dealing with large error margins.
Operate in Unstructured Environments: The ability to function under one-sided assumptions allows AI systems to integrate and compare data from external sources that are not guaranteed to be perfectly structured or separated, making it ideal for decentralized data collection and comparison tasks.

Abstract

Fuzzy private set intersection (PSI) enables two parties to identify approximately matching elements between their input sets, where two elements are considered a match if their distance is at most a threshold δ under a given metric. Although substantial progress has been made, existing constructions for general Minkowski distances either rely on strong two-sided geometric separation assumptions or incur substantial overhead under one-sided assumptions. In this work, we present the first concretely efficient fuzzy PSI protocols for general L p in[1, infinity] distances under one-sided assumptions, relying solely on lightweight symmetric-key primitives. Our constructions support both sender-sided and receiver-sided settings. We further study sparser input distributions and present more efficient protocols tailored to this case. To reduce the overhead scaling with δ, we non-trivially incorporate prefix trie techniques into our protocols, achieving O(δ) complexity for general L p in[1, infinity] distances for the first time, improving upon O((δ) d) or O(δ) complexities of prior works. Extensive experiments, across a wide range of parameter settings, show that our protocols significantly outperform prior works under the same assumptions. Specifically, against van Baarsen and Pu (EUROCRYPT'24), our protocols achieve up to 239 times faster computation and up to 20 times lower communication. Against Dang et al. (CCS'25), we achieve up to 518 times speedup and up to 63 times communication reduction. Against Bui et al. (ASIACRYPT'25), we achieve up to 4818 times faster computation and up to 282 times lower communication.