Title: Dual-Target Attacks Uncover Flaws in Time Series Classifier Explanation Reliability

Abstract: The robustness of interpretable deep learning systems for time series analysis is frequently evaluated based on the temporal consistency of their explanations, a metric often assumed to guarantee reliability. This study demonstrates that such an assumption is flawed: it is possible to adversarially decouple predictions from their corresponding explanations. This decoupling allows for precise, targeted misclassifications while the generated explanations remain plausible and aligned with a specific reference rationale. To address this, we introduce TSEF (Time Series Explanation Fooler), a dual-target attack mechanism that simultaneously manipulates the outputs of both the classifier and the explainer. Unlike traditional single-objective attacks that cause misclassification by disrupting explanations and diffusing attribution mass widely, TSEF secures targeted shifts in prediction outcomes while maintaining explanation consistency with the reference. Our experiments across various datasets and explainer architectures consistently show that explanation stability is an unreliable indicator of decision-making robustness. These findings highlight the necessity for coupling-aware robustness assessments to ensure the trustworthiness of time series applications.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC