Remix-Proof Retrieval
收藏资源简介:
为了支持我们的版权侵权检测方法,我们开发了一个专门用于基于检索的匹配的结构化数据集。数据集来源于Hugging Face上的`benjamin-paine/free-music-archive-small`集合,共包含7,916个音频轨道。每个源歌曲(我们视为受版权保护的歌曲)的持续时间约为30秒,并均匀分布在8种流派中,确保音乐风格的多样性。为了模拟真实的版权侵权场景,我们从每个轨道中随机提取5秒的片段作为查询。这种方法反映了现实世界中可能只有音频片段可用于比较的情况。为了测试歌曲识别对抗规避策略的能力,我们进一步增强了这些五秒片段作为受版权保护歌曲的伪装副本。
To support our copyright infringement detection method, we developed a structured dataset specifically designed for retrieval-based matching. The dataset is sourced from the `benjamin-paine/free-music-archive-small` collection on Hugging Face, and it contains a total of 7,916 audio tracks. Each source song (which we treat as copyrighted work) has a duration of approximately 30 seconds, and they are evenly distributed across 8 genres to ensure diversity in musical styles. To simulate real-world copyright infringement scenarios, we randomly extracted 5-second clips from each track as queries. This approach reflects the real-world situation where only audio clips may be available for comparison. To test the capability of song recognition against adversarial evasion strategies, we further enhanced these 5-second clips to act as disguised copies of the copyrighted songs.
数据集概述
基本信息
- 数据集名称: Remix-Proof Retrieval
- 用途: 检测数字音乐中的版权侵权行为,支持音乐信息检索(MIR)任务
- 数据集地址: HuggingFace
数据来源
- 原始数据集:
benjamin-paine/free-music-archive-small(来自Hugging Face) - 音频数量: 7,916条音轨
- 音频时长: 每条音轨约30秒
- 音乐风格: 均匀分布在8种不同流派中
数据构造
- 查询片段: 从每条音轨中随机提取5秒片段作为查询
- 模拟侵权场景: 对5秒片段进行多种音频增强处理,生成伪装副本
关键特性
- 针对版权侵权检测的鲁棒音频编码
- 可抵抗常见音频扰动的微调模型
- 音乐信息检索评估框架
- 对抗性增强的基准测试
- 对比学习技术实现
贡献者
- Abhir Karande (akarande@usc.edu)
- Ayush Goyal (ayushgoy@usc.edu)
- Harrison Pearl (hpearl@usc.edu)
- Matthew Hong (hongmm@usc.edu)
- Ryan Lee (ryantlee@usc.edu)
- Spencer Cobb (srcobb@usc.edu)
- Yi-Chieh Chiu (ychiu443@usc.edu)
引用信息
@misc{lee2025remixproof, title={Remix-Proof Retrieval: Robust Audio Encoders for Detecting Copyright Infringement}, author={Lee, Ryan* and Chiu, Yi-Chieh* and Karande, Abhir* and Goyal, Ayush and Pearl, Harrison and Hong, Matthew and Cobb, Spencer}, year={2025}, note={*Equal contribution}, publisher={GitHub}, howpublished={url{https://github.com/username/remix-proof-retrieval}} }




