一种听觉感知引导的翻唱歌曲识别系统对抗样本生成方法
An Auditory Perception-Guided Adversarial Example Generation Method for Cover Song Identification Systems
摘要
本文针对翻唱歌曲识别(CSI)任务中的关键缺陷展开研究.现有 CSI 模型在特征空间分布中未能充分考虑听觉感知关系,导致模型性能受限且存在安全风险.为验证该问题,本文引入听觉感知模型,提出一种面向翻唱歌曲识别系统的对抗样本生成方法.该方法基于音乐特征分布生成对抗噪声片段,并将其注入待识别音乐,在尽量保持人类听感一致的同时,诱导目标模型产生误识别.实验结果表明,该方法能够有效生成针对指定 CSI 模型的黑盒对抗样本,攻击成功率达 86.4%,可为提升翻唱歌曲识别系统的鲁棒性与安全性提供参考.
This paper investigates a critical deficiency in Cover Song Identification(CSI)tasks.Existing CSI models fail to adequately account for auditory perception relationships in feature space distribution,which limits model performance and introduces security risks.To address this issue,we introduce an auditory perception model and propose an adversarial example generation method for CSI systems.The method generates adversarial noise segments based on the distribution of musical features and injects them into the music to be identified,thereby inducing misidentification by the target model while preserving human auditory perception as much as possible.Experimental results demonstrate that the proposed method can effectively generate black-box adversarial examples against specified CSI models,achieving an attack success rate of 86.4%.This work provides a reference for improving the robustness and security of cover song identification systems.
年,卷(期):
2026, 65(5)
分类号:
TP181
机标分类号:
TP393.1
TN97
E843
在线出版日期:
2026年11月4日
基金项目:
RDF-22-02-046:西交利物浦大学科研发展基金,