Scholarlink

Time-series data has been increasingly utilized in critical domains making it vulnerable to stealthy adversarial attacks. Explainable AI (XAI) has made major enhancements to model transparency specifically in non-sequential domains (Image, NLP and tabular) but is less researched in time series data due to complexity of temporal dependency in sequential data. This paper systematically reviews around 60 studies published up to June 2025, examining XAI-based approaches for time series, as well as recent XAI extensions tailored for time series data. Various attack types are discussed in terms of their effectiveness, and the performance of recent detection frameworks against these attacks is analyzed, with particular focus on XAI-integrated systems, including SHAP, LIME, LRP, and formal verifications. Despite their usage, most XAI-enhanced approaches, especially those relying on SHAP, face challenges in efficiency and risk explanation instability under noise, with limited evaluation against subtle partial adversarial attacks. While LRP offers computational efficiency vital for high stakes domain real-time detection, its architectural dependency restricts its adaptability across diverse models. This review highlights major unsolved limitations and concludes with prioritizing the need for efficient, robust, temporally aware, and architecture-agnostic frameworks, possibly extending LRP, tailored specifically for time series data.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Explainable AI for Time Series: A Systematic Review of Robust and Efficient Detection of Stealthy Adversarial Attacks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mohammed Hashim Kalam,
  • Guhanathan Poravi

摘要

Time-series data has been increasingly utilized in critical domains making it vulnerable to stealthy adversarial attacks. Explainable AI (XAI) has made major enhancements to model transparency specifically in non-sequential domains (Image, NLP and tabular) but is less researched in time series data due to complexity of temporal dependency in sequential data. This paper systematically reviews around 60 studies published up to June 2025, examining XAI-based approaches for time series, as well as recent XAI extensions tailored for time series data. Various attack types are discussed in terms of their effectiveness, and the performance of recent detection frameworks against these attacks is analyzed, with particular focus on XAI-integrated systems, including SHAP, LIME, LRP, and formal verifications. Despite their usage, most XAI-enhanced approaches, especially those relying on SHAP, face challenges in efficiency and risk explanation instability under noise, with limited evaluation against subtle partial adversarial attacks. While LRP offers computational efficiency vital for high stakes domain real-time detection, its architectural dependency restricts its adaptability across diverse models. This review highlights major unsolved limitations and concludes with prioritizing the need for efficient, robust, temporally aware, and architecture-agnostic frameworks, possibly extending LRP, tailored specifically for time series data.