Personalized w-Event Privacy for Infinite Stream Estimation

arXiv:2605.09054 · cs.DB, cs.CR, cs.IR · Submitted 2026-08-20 · 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 "Personalized w-Event Privacy for Infinite Stream Estimation".

Jane: The paper was written by Leilei Du, Xu Zhou, Peng Cheng, Lei Chen, Xuemin Lin et al. from Hunan University and Tongji University and Hong Kong University of Science and Technology (Guangzhou) and Hong Kong University of Science and Technology and Shanghai Jiaotong University and Xi’an Jiaotong University.

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.

Title: Tom: We have a massive one to start with today, Jane.

Jane: You're talking about "Personalized w-Event Privacy for Infinite Stream Estimation," aren't you?

Tom: That's the one!

Jane: The title sounds like a mouthful, but it's actually quite beautiful once you peel back the layers.

Tom: It's a collaborative effort from a huge team at places like Hunan University and HKUST.

Jane: I noticed Leilei Du and Kenli Li are leading the charge on this.

Tom: They've pulled in experts from Tongji and Shanghai Jiaotong too.

Jane: It makes sense because this problem is huge.

Lu: It's more than just huge, Tom; it's a complete shift in how we think about data ownership.

Tom: How so, Lu?

Lu: Most privacy research treats everyone as a single, identical block of data.

Jane: Which is exactly what this paper is pushing back against.

Tom: I was reading the intro, and it's all about moving away from that "one-size-fits-all" approach.

Meng: I wonder if they've actually thought about the computational overhead of doing that for millions of users.

Jane: That's a fair question, Meng, because calculating different rules for everyone sounds like a nightmare.

Tom: The authors seem to have a plan for that.

Meng: I'll believe it when I see the implementation details.

Lalam: The real magic is in the cultural shift this enables.

Jane: You mean the way it treats people as individuals?

Lalam: Exactly, because it acknowledges that a celebrity's privacy needs are different from a student's.

Tom: That's a great way to put it.

Jane: We should probably get into what they are actually doing with these "w-events" to see how it works.

Summary: Jane: So, to understand this paper, we have to look at how "w-event privacy" works.

Tom: It's basically saying your data is protected within a certain window of time, right?

Jane: Yes, if your window is eight, your information stays private across eight consecutive events.

Tom: But the problem is that the current systems force everyone into the same window.

Jane: They used a car-hailing example to show how messy that gets.

Tom: I loved that example where you have a hundred drivers.

Jane: Some drivers only need a tiny bit of protection, while others need a much larger window.

Tom: If you set the window to the largest requirement, you end up wasting privacy budget on everyone else.

Jane: And that waste leads to a lot of extra noise in the data.

Lu: It's like trying to protect a butterfly and an elephant with the same size cage.

Tom: That's a vivid image, Lu.

Lu: The paper is trying to solve the "infinite stream" part, where data never stops flowing.

Meng: How do they handle the fact that these users aren't just different, but they're changing?

Jane: That's the "heterogeneous" part they mention.

Tom: They're talking about users having different privacy budgets and different window sizes all at once.

Meng: Managing that many moving parts in a live stream sounds incredibly difficult.

Jane: It is, which is why they focus on making the aggregate results accurate for everyone.

Tom: They want to publish one single statistic that respects everyone's unique rules.

Lu: It's a massive balancing act between being useful and being private.

Lalam: It's also about giving people the power to decide their own boundaries.

Jane: Which leads us directly into the specific mechanisms they actually built to make this happen.

Improvements: Tom: They didn't just point out the problem; they actually proposed several new mechanisms.

Jane: They started with the Personalized Window Size Mechanism, or PWSM.

Tom: Which then branches out into two main strategies: PBD and PBA.

Jane: PBD is about distributing the budget, while PBA is about absorbing it.

Tom: I found the "absorption" idea really clever.

Jane: It lets a user borrow privacy budget from future time slots to make the current one more accurate.

Tom: And then they took it even further with the dynamic versions, DPBD and DPBA.

Jane: Those handle people who change their minds about privacy settings over time.

Lu: The way they handle "backward" and "forward" windows is brilliant.

Tom: Wait, explain that, Lu.

Lu: A backward window looks at your recent history, while a forward window protects your upcoming moves.

Meng: I'm looking at these error rates in the results section.

Jane: What caught your eye, Meng?

Meng: The DPBD method reduces the average error by at least sixty-two point seven percent compared to the old way.

Tom: And DPBA beats the other method by fifty-three point six percent.

Meng: Those are significant jumps for a real-world system.

Jane: It shows that being personalized actually makes the data more useful, not less.

Lu: It proves that we don't have to sacrifice accuracy to respect individual choices.

Lalam: It's a vision of a future where technology adapts to human nuance.

Tom: We've covered a lot of ground, so let's wrap this up.

Conclusion: Tom: This has been a heavy one, but such a rewarding discussion.

Jane: We've seen how "Personalized w-Event Privacy for Infinite Stream Estimation" changes the game.

Tom: It moves us from a rigid, uniform system to one that actually understands individual needs.

Jane: It's a massive step forward for anyone working with real-time data.

Lu: I think this opens the door for AI that respects the fluidity of human life.

Meng: From my side, I'm looking forward to seeing how these algorithms scale in production.

Lalam: It's a beautiful step toward a more respectful digital culture.

Tom: Thanks to everyone for joining us.

Jane: We'll see you next time!

Leilei Du, Xu Zhou, Peng Cheng, Lei Chen, Xuemin Lin, Wei Xi, Kenli Li

Hunan University · Tongji University · Hong Kong University of Science and Technology (Guangzhou) · Hong Kong University of Science and Technology · Shanghai Jiaotong University · Xi’an Jiaotong University

cs.DB, cs.CR, cs.IR

Submitted: 2026-08-20

Updated: 2026-08-21

Code: https://github.com/dulei715/DynamicWEventCode

Importance score: 79/100

The gist: This paper addresses the limitations of existing w-event privacy mechanisms for infinite data streams, noting that "existing w-event privacy studies on infinite data stream typically focus only on

Key concepts

w-Event Privacy: A method of protecting data by ensuring that information remains private within a specific window of time (w). The length of this window determines how many consecutive events must pass before the data is considered unprotected.
Personalized Approach: Moving away from treating all users as identical blocks of data. This approach acknowledges that different individuals, such as celebrities versus students, have unique and varying privacy needs and rules.
Infinite Stream Estimation: The challenge of analyzing data that never stops flowing (a live stream). The paper addresses how to maintain accuracy and privacy when dealing with continuous, unending data inputs.
PWSM/DPBD/DPBA: Mechanisms proposed by the authors to manage personalized privacy. PWSM is the core mechanism, while DPBD and DPBA are dynamic versions that allow users to change their privacy settings over time.

Terminology

Differential Privacy — 61× in this paper · explained in 4 episodes
Budget Absorption — 20× in this paper
Budget Distribution — 18× in this paper
Accuracy — 17× in this paper · explained in 1 episode
Personalized Budget Distribution — 15× in this paper
Personalized Budget Absorption — 12× in this paper
Personalized Window Size Mechanism — 11× in this paper
Dissimilarity Calculation — 9× in this paper
Electronic Supplementary Material — 9× in this paper
Mean Relative Error — 8× in this paper
Average Mean Relative Error — 6× in this paper
Dynamic Personalized Budget Absorption — 6× in this paper
Dynamic Personalized Budget Distribution — 6× in this paper
Heterogeneous Differential Privacy — 4× in this paper
Sampling Mechanism — 4× in this paper
Core — 3× in this paper · explained in 2 episodes
Differential Privacy (DP) — 3× in this paper · explained in 7 episodes
Differentially Private — 3× in this paper · explained in 3 episodes
Dynamic Personalized Window Size Mechanism — 3× in this paper
Optimal Budget Selection — 3× in this paper
Average Jensen-Shannon Divergence — 2× in this paper
Average Wasserstein Distance — 2× in this paper
Error Decomposition — 2× in this paper · explained in 3 episodes
One-sided Differential Privacy — 2× in this paper
Sparsity — 2× in this paper · explained in 4 episodes
Unified Framework — 2× in this paper · explained in 5 episodes
Wasserstein Distance — 2× in this paper · explained in 4 episodes
Adaptability — 1× in this paper · explained in 1 episode
Adaptation — 1× in this paper · explained in 1 episode
Adaptive Sampling — 1× in this paper · explained in 1 episode
Adaptive Strategy — 1× in this paper · explained in 1 episode
Computational overhead — 1× in this paper · explained in 2 episodes
Consistency — 1× in this paper · explained in 3 episodes
Data Heterogeneity — 1× in this paper · explained in 3 episodes
Generalization — 1× in this paper · explained in 13 episodes
Personalized LDP Budget Uniform — 1× in this paper
Recursive Binary Tree Mechanism — 1× in this paper
Robustness — 1× in this paper · explained in 5 episodes

Summary

This paper addresses the limitations of existing w-event privacy mechanisms for infinite data streams, noting that existing w-event privacy studies on infinite data stream typically focus only on homogeneous privacy requirements for all users, which provides inadequate privacy protection for some users while unnecessarily increasing data error rates (excessive privacy protection) for others. To resolve this, the authors propose personalized w-event privacy protection that enables users to set different privacy requirements in private data stream estimation.

The research introduces a unified framework for personalized stream release. In the fixed setting, where users maintain constant requirements, the authors design a Personalized Window Size Mechanism (PWSM) that allows users to maintain personalized privacy requirements at each time slot. This framework includes two specific solutions: Personalized Budget Distribution (PBD), which ensures that the privacy budget for the next time step is at least equal to the amount consumed in the previous release, and Personalized Budget Absorption (PBA), which enhances the current time slot’s privacy budget by combining the privacy budget from the previous k time slots and borrowing from the next k time slots. The authors note that PBD is preferable when the stream exhibits persistent rapid changes, whereas PBA is more suitable for relatively smooth streams.

The paper extends this to a more general problem termed Dynamic Personalized w-Event Private Publishing for Infinite Data Streams (DPWEPP-IDS), where each user may specify time-varying backward and forward privacy requirements. This dynamic setting is characterized by the notation (tau, w B, w F, E B, E F) -Event (E B, E F) -Personalized Differential Privacy ((tau, w B, w F, E B, E F)-EPDP). In this model, a "backward requirement is anchored at the current time slot and constrains privacy loss accumulated from the recent past, whereas a forward requirement is declared at the current time slot and constrains feasible future releases in the upcoming window. To address this, the authors generalize the framework to the Dynamic Personalized Window Size Mechanism (DPWSM), featuring two mechanisms: Dynamic Personalized Budget Distribution (DPBD) and Dynamic Personalized Budget Absorption (DPBA)."

The technical core of these mechanisms relies on Optimal Budget Selection (OBS), which is used to determine a release threshold under heterogeneous privacy budgets, and the Sampling Mechanism (SM), which is used to achieve epsilon-PDP. The system utilizes a Dissimilarity Calculation (DC) to perform personalized dissimilarity estimation and adaptive release, deciding whether to publish a new obfuscated estimation or skip (i.e., use the last published one) by comparing dis to sqrt err.

The authors provide formal proofs for privacy guarantees and establish error upper bounds for each method. Complexity analysis shows that the fixed mechanisms have a memory complexity O(n times w max) and time complexities... are both O(n), while the dynamic mechanisms have a time complexity of O(w max + w max times n) and a memory complexity of O(w max + w max times n).

Experimental results on real datasets (Taxi and Foursquare) and synthetic datasets (TLNS, Sin, and Log) demonstrate significant utility improvements. Specifically, for real datasets, DPBD reduces AMRE by at least 62.7% compared with BD, and for synthetic datasets, DPBA reduces AMRE by at least 53.6% compared with BA. The results confirm that the proposed mechanisms improve utility over classical homogeneous w-event baselines in the heterogeneous personalized setting.

Improvements for AI systems

1. Implementation of Heterogeneous Real-Time Stream Aggregators (e.g., Smart City Traffic or Energy Grid Monitoring)

Improvement: Integrate the Personalized Window Size Mechanism (PWSM), specifically utilizing Personalized Budget Distribution (PBD) for high-volatility data and Personalized Budget Absorption (PBA) for stable data.
Capability: The AI system can ingest data from millions of IoT devices where each device has unique, non-uniform privacy requirements (e.g., a high-security government vehicle requiring a long privacy window w and small budget epsilon, versus a standard consumer vehicle requiring a short window). The system will automatically switch between PBD and PBA to minimize Average Mean Relative Error (AMRE) based on the stream's temporal characteristics, ensuring high-accuracy aggregate statistics (like traffic flow or load demand) without violating individual privacy tiers.

2. Development of Context-Aware Federated Learning (FL) Orchestrators

Improvement: Incorporate the Dynamic Personalized Window Size Mechanism (DPWSM) using DPBD and DPBA to manage client updates.
Capability: The orchestrator can handle mobile or wearable device clients whose privacy preferences change dynamically based on context (e.g., a user requesting high privacy during sleep/sensitive hours and lower privacy during active hours). The system can reconcile these heterogeneous, time-varying backward and forward privacy requirements into a single, valid system-level release decision, maintaining global model convergence and utility while strictly adhering to the user's evolving privacy windows.

3. Deployment of Adaptive Privacy-Preserving Anomaly Detection Systems (Cybersecurity/Log Analysis)

Improvement: Implement the Optimal Budget Selection (OBS) and Sampling Mechanism (SM) framework within the detection pipeline.
Capability: The AI system can perform real-time statistical monitoring on infinite data streams with varying levels of sensitivity. By utilizing OBS, the system will automatically calculate the optimal privacy budget threshold (epsilon theta) to minimize the combined impact of sampling variance and Laplace noise. This allows the system to maintain high-fidelity detection of rapid, abrupt anomalies (using PBD) while providing highly accurate, low-noise monitoring during steady-state operations (using PBA).

4. Construction of Privacy-Preserving Personalized Recommendation Engines

Improvement: Apply the (τ, w B, w F)-Event (E B, E F)-Personalized Differential Privacy (EPDP) framework to user preference aggregation.
Capability: The system can aggregate real-time user interest streams where users can opt-in to higher accuracy (lower privacy) for specific, time-limited windows (e.g., during a holiday shopping season) and opt-out (higher privacy) during other periods. The system will manage these dynamic forward and backward constraints to provide highly relevant, real-time trending recommendations while guaranteeing that no user's specific temporal behavior is leaked beyond their declared privacy window.

Sources

Differential Privacy in the Shuffle Model: A Survey of Separations