IU: Imperceptible Universal Backdoor Attack

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Lin, Hsin, Chen, Yan-Lun, Hwang, Ren-Hung, Yu, Chia-Mu
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911475457589248
author Lin, Hsin
Chen, Yan-Lun
Hwang, Ren-Hung
Yu, Chia-Mu
author_facet Lin, Hsin
Chen, Yan-Lun
Hwang, Ren-Hung
Yu, Chia-Mu
contents Backdoor attacks pose a critical threat to the security of deep neural networks, yet existing efforts on universal backdoors often rely on visually salient patterns, making them easier to detect and less practical at scale. In this work, we introduce a novel imperceptible universal backdoor attack that simultaneously controls all target classes with minimal poisoning while preserving stealth. Our key idea is to leverage graph convolutional networks (GCNs) to model inter-class relationships and generate class-specific perturbations that are both effective and visually invisible. The proposed framework optimizes a dual-objective loss that balances stealthiness (measured by perceptual similarity metrics such as PSNR) and attack success rate (ASR), enabling scalable, multi-target backdoor injection. Extensive experiments on ImageNet-1K with ResNet architectures demonstrate that our method achieves high ASR (up to 91.3%) under poisoning rates as low as 0.16%, while maintaining benign accuracy and evading state-of-the-art defenses. These results highlight the emerging risks of invisible universal backdoors and call for more robust detection and mitigation strategies.
format Preprint
id arxiv_https___arxiv_org_abs_2603_00711
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle IU: Imperceptible Universal Backdoor Attack
Lin, Hsin
Chen, Yan-Lun
Hwang, Ren-Hung
Yu, Chia-Mu
Cryptography and Security
Computer Vision and Pattern Recognition
Machine Learning
Backdoor attacks pose a critical threat to the security of deep neural networks, yet existing efforts on universal backdoors often rely on visually salient patterns, making them easier to detect and less practical at scale. In this work, we introduce a novel imperceptible universal backdoor attack that simultaneously controls all target classes with minimal poisoning while preserving stealth. Our key idea is to leverage graph convolutional networks (GCNs) to model inter-class relationships and generate class-specific perturbations that are both effective and visually invisible. The proposed framework optimizes a dual-objective loss that balances stealthiness (measured by perceptual similarity metrics such as PSNR) and attack success rate (ASR), enabling scalable, multi-target backdoor injection. Extensive experiments on ImageNet-1K with ResNet architectures demonstrate that our method achieves high ASR (up to 91.3%) under poisoning rates as low as 0.16%, while maintaining benign accuracy and evading state-of-the-art defenses. These results highlight the emerging risks of invisible universal backdoors and call for more robust detection and mitigation strategies.
title IU: Imperceptible Universal Backdoor Attack
topic Cryptography and Security
Computer Vision and Pattern Recognition
Machine Learning
url https://arxiv.org/abs/2603.00711