Context-Aware Functional Modeling for Android Third-Party Library Detection
原文链接: https://arxiv.org/abs/2609.31409
作者: Dihao Fan et al.
发布时间: 2026-09-25
源: arxiv
摘要
LibFan(arXiv 2609.31409)是一种基于学习的 Android 第三方库检测方法,核心是上下文感知的功能建模,由方法级对比学习(含 outgoing call + 类上下文)与库级功能分区两部分组成。方法级表征对混淆、shrink、优化更鲁棒,库级功能分区支持部分复用。自建 200 app + 46 漏洞 TPL 基准、四种变换配置;R8 full mode 下库级 F1 81.3%(SOTA 相对 +64.9%)、版本级 47.6%(+35.6%)。R8 full mode 是最贴近现行 release 构建的设定,数字可信度更高。
English Summary
LibFan (arXiv 2609.31409) is a learning-based Android third-party library detection approach built on context-aware functional modeling, with two complementary components: context-aware contrastive learning at the method level (incorporating outgoing calls and class-level context) and functional partitioning at the library level. Method-level representations harden the detector against obfuscation, shrinking, and optimization, while library-level partitioning supports partial reuse. On a self-built benchmark of 200 apps and 46 vulnerable TPLs across four transformation configurations, LibFan reaches library-level F1 of 81.3% under R8 full mode (+64.9% over prior SOTA) and version-level F1 of 47.6% (+35.6%). The R8 full-mode result is the most production-realistic number in the table.
为什么值得关注
主题线扩展 AAIF android/third-party-library/tpl-detection/contrastive-learning 等主题。