Black-Box Adversarial Attack on Time Series Classification

With the increasing use of deep neural network (DNN) in time series classification (TSC), recent work reveals threat adversarial attack, where adversary can construct examples to cause model mistakes. However, existing researches on attack TSC typically adopt an unrealistic white-box setting with details transparent adversary. In this work, we study a more rigorous black-box detection applied, which restricts gradient access and requires example be also stealthy. Theoretical analyses reveal that key lies in: estimating diversity non-convexity models resolved, restricting l0 norm perturbation samples. Towards end, propose new framework named BlackTreeS, solves hard optimization issue for construction two simple yet effective modules. particular, tree search strategy find influential positions sequence, independently estimate gradients these positions. Extensive experiments three real-world datasets five DNN based validate effectiveness e.g., it improves success rate from 19.3% 27.3%, decreases 90.9% 6.8% LSTM UWave dataset.