Local adversarial attack of time series forecasting based on mutual information

In recent years, as time series forecasting has been increasingly integrated into daily operations, scholarly attention has pivoted towards its security and fidelity. Traditional deep learning models for time series prediction often exhibit a lack of adversarial robustness, rendering them vulnerable to strategically crafted perturbations that may skew predictive outcomes significantly. Presently, adversarial interventions in time series data predominantly leverage global perturbations; however, such manipulations are highly conspicuous and thus easily detectable given the intrinsic sensitivity of time series data to disturbances. In response, we introduce a novel gradient-free, sparse black-box attack methodology, denoted as kMGAT, which harnesses mutual information and genetic algorithms to adeptly balance the stealth and efficacy of adversarial sample generation. This approach is capable of orchestrating potent attacks without materially distorting the original data corpus. Rigorous comparative evaluations across diverse deep learning architectures have demonstrated that kMGAT can fabricate adversarial samples that are both challenging to detect and capable of significantly altering the predictive patterns of the models. This methodology has yielded commendable results in several demanding applications, including forecasting the occupancy rates of hospital parking, electrical consumption metrics, and subway traffic volumes.