Go to ICLR 2025 Conference homepageGradient-Free Adversarial Attack on Time Series Regression: Targeting XAI Explanations
Explainable Artificial Intelligence (XAI) sheds light on the decision-making ground of black-box models by offering explanations. These explanations need to be robust for trustworthy time series regression applications in high-stake areas like medicine or finance, which yet remains largely unexplored. Furthermore, most adversarial attack methods currently rely on white-box strategies, which require access to gradient information from both the model and the XAI method. In real-world scenarios, such information is often difficult or impossible to obtain. To address these challenges, we propose a novel gradient-free adversarial attack method specifically designed for time series explanations, targeting non-differentiable XAI techniques. To enhance the effectiveness of our method for time series data, we introduce an attack objective function based on Dynamic Time Warping (DTW). Additionally, we implement an explanation-based local attack strategy, which ensures that the adversarial perturbations remain imperceptible within the time series data. In our experiments, we generate adversarial examples to attack four different XAI methods across three black-box models, using two time series datasets. The results reveal the vulnerability of current non-differentiable XAI methods. Furthermore, by comparing our approach with existing attack methods, we demonstrate the superiority of our proposed objective function and local attack strategy.