From Adversity to Advantage: Diffusion Models for Improved Detection Under Attack

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Adversarial patch attacks threaten the reliability of object detectors by causing severe misclassifications, especially in safety-critical environments. In this work, we propose a comprehensive defense pipeline that not only restores detection performance but also significantly improves it. Our method leverages a latent diffusion model to recover semantically coherent regions affected by adversarial patches, leading to confidence gains of +26.61% for YOLOv5 and +26.91% for YOLOv7 - exceeding the models’ original predictions on clean images. In contrast to prior approaches that merely attempt restoration, we demonstrate that diffusion models can enhance object detection performance under attack, while maintaining practical efficiency with of inference time. We also present an optimized attack strategy based on EigenCAM and grid search, which identifies and targets the most vulnerable regions of the image. Experimental results show that our method consistently outperforms classical and recent defenses such as JPEG compression, spatial smoothing, SAC [17], and DIFFender [8], both in robustness and in detection confidence recovery. These findings highlight the potential of generative models not only for defense but for strengthening object detectors in adversarial scenarios.

Original languageEnglish
Title of host publicationCyber Security, Cryptology, and Machine Learning - 9th International Symposium, CSCML 2025, Proceedings
EditorsAdi Akavia, Shlomi Dolev, Anna Lysyanskaya, Rami Puzis
PublisherSpringer Science and Business Media Deutschland GmbH
Pages104-121
Number of pages18
ISBN (Print)9783032107589
DOIs
StatePublished - 1 Jan 2026
Event9th International Symposium on Cyber Security, Cryptology, and Machine Learning, CSCML 2025 - Be'er Sheva, Israel
Duration: 4 Dec 20255 Dec 2025

Publication series

NameLecture Notes in Computer Science
Volume16244 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference9th International Symposium on Cyber Security, Cryptology, and Machine Learning, CSCML 2025
Country/TerritoryIsrael
CityBe'er Sheva
Period4/12/255/12/25

Keywords

  • adversarial attacks
  • adversarial patch defense
  • detection confidence
  • object detection

ASJC Scopus subject areas

  • Theoretical Computer Science
  • General Computer Science

Fingerprint

Dive into the research topics of 'From Adversity to Advantage: Diffusion Models for Improved Detection Under Attack'. Together they form a unique fingerprint.

Cite this