I Can Still Steal Your Encoder: A Defense-Penetrating Encoder-Stealing Attack

Rongbin Xiao, Changyu Dong, Jie Zhang, Yan Pang, Zihan Xie, Han Wu

Research output: Chapter or section in a book/report/conference proceedingChapter in a published conference proceeding

Abstract

The rise of Encoder-as-a-Service (EaaS) has made pre-trained encoders accessible for various AI tasks, but this has introduced significant security concerns, particularly with model stealing attacks. While defenses like the B4B mechanism [6] have been proposed to protect against such attacks, we reveal critical vulnerabilities in B4B’s strategies. B4B employs techniques such as embedding space coverage estimation, cost-based perturbation, and embedding transformations to thwart attackers. However, we introduce the first defense-penetrating attack that bypasses these protections. Our attack effectively circumvents all three defense mechanisms, enabling attackers to steal high-quality encoders with minimal degradation in performance. Extensive experiments show that the stolen encoder performs almost as well as the original, highlighting the weaknesses in B4B and similar defenses. Our work exposes significant gaps in the security of EaaS systems and calls for more robust, active defense strategies against model stealing.

Original languageEnglish
Title of host publicationPattern Recognition and Computer Vision - 8th Chinese Conference, PRCV 2025, Proceedings
EditorsJosef Kittler, Hongkai Xiong, Weiyao Lin, Jian Yang, Xilin Chen, Jiwen Lu, Jingyi Yu, Weishi Zheng
Place of PublicationSingapore
PublisherSpringer
Pages483-501
Number of pages19
ISBN (Electronic)9789819557646
ISBN (Print)9789819557639
DOIs
Publication statusPublished - 24 Jan 2026
Event8th Chinese Conference on Pattern Recognition and Computer Vision, PRCV 2025 - Shanghai, China
Duration: 15 Oct 202518 Oct 2025

Publication series

NameLecture Notes in Computer Science
Volume16289 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference8th Chinese Conference on Pattern Recognition and Computer Vision, PRCV 2025
Country/TerritoryChina
CityShanghai
Period15/10/2518/10/25

Keywords

  • Defense-Penetrating Attack
  • Model Stealing
  • Representation Learning

ASJC Scopus subject areas

  • Theoretical Computer Science
  • General Computer Science

Fingerprint

Dive into the research topics of 'I Can Still Steal Your Encoder: A Defense-Penetrating Encoder-Stealing Attack'. Together they form a unique fingerprint.

Cite this