Skip to main navigation Skip to search Skip to main content

Dynamic Privacy Budget Allocation Mechanism Based on Multi-Dimensional Query Management

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

Abstract

This study proposes a dynamic adjustment mechanism for evaluating privacy budget parameters to address privacy protection challenges in multi-turn interactions of generative chatbots. By classifying queries across multiple dimensions - including domain, sensitivity, and risk - we construct a comprehensive privacy-preserving framework. Traditional differentially private training methods typically allocate a fixed privacy budget under the assumption of static query behaviors, which often fails to reflect the dynamic nature of real-world interactions. To overcome this limitation, we introduce a query-adaptive mechanism that adjusts the privacy budget per query based on characteristics such as sensitivity, frequency, and temporal patterns. This approach enables more precise control of cumulative privacy loss and mitigates the risks of repeated or excessive querying. Simulation results show that our method achieves an effective balance between privacy preservation and response quality across various query frequencies, domains, and scenarios. This study offers an innovative solution for privacy-preserving generative chatbots and contributes to both the theoretical and practical development of differential privacy in interactive applications.

Original languageEnglish
Title of host publicationInternational Conference on Electrical, Computer, and Energy Technologies, ICECET 2025
PublisherInstitute of Electrical and Electronics Engineers Inc.
ISBN (Electronic)9798331535599
DOIs
StatePublished - 2025
EventIEEE International Conference on Electrical, Computer and Energy Technologies, ICECET 2025 - Paris, France
Duration: 03 07 2025 → 06 07 2025

Publication series

NameInternational Conference on Electrical, Computer, and Energy Technologies, ICECET 2025

Conference

ConferenceIEEE International Conference on Electrical, Computer and Energy Technologies, ICECET 2025
Country/TerritoryFrance
CityParis
Period03/07/25 → 06/07/25

Bibliographical note

Publisher Copyright:
© 2025 IEEE.

Keywords

  • Adaptive Privacy Protection
  • Differential Privacy
  • Dynamic Privacy Budget
  • Generative Chatbot

Fingerprint

Dive into the research topics of 'Dynamic Privacy Budget Allocation Mechanism Based on Multi-Dimensional Query Management'. Together they form a unique fingerprint.

Cite this