ASAT: Adaptively Scaled Adversarial Training in Time Series

Abstract

Adversarial training is a method for enhancing neural networks to improve the robustness against adversarial examples. Besides the security concerns of potential adversarial examples, adversarial training can also improve the performance of the neural networks, train robust neural networks, and provide interpretability for neural networks. In this work, we take the first step to introduce adversarial training in time series analysis by taking the finance field as an example. Rethinking existing researches of adversarial training, we propose the adaptively scaled adversarial training (ASAT) in time series analysis, by treating data at different time slots with timedependent importance weights. Experimental results show that the proposed ASAT can improve both the accuracy and the adversarial robustness of neural networks. Besides enhancing neural networks, we also propose the dimension-wise adversarial sensitivity indicator to probe the sensitivities and importance of input dimensions. With the proposed indicator, we can explain the decision bases of black box neural networks. CCS CONCEPTS• Computing methodologies → Neural networks.

Distributional Correlation--Aware Knowledge Distillation for Stock Trading Volume Prediction

“…y T is adopted as the prediction. We set ρ = 0.04 following [32]. Teacher-free methods, which requires no teacher model for training the student.…”

“…Besides, temporal mixture ensemble models [1], Bayesian auto-regressive models [12] and graph neural networks [33] are also explored in volume prediction. [32] train a Transformer model [30] with adversarial objectives to improve the model performance and robustness at the same time. In this paper, we focus on distilling a more efficient trading volume prediction model and adopt the powerful Transformer as the backbone model for distillation.…”

“…Large deep neural networks (DNNs) like Transformer [30] have achieved superior performance in various areas like computer vision [6], natural language processing [5] and time series forecasting problems like stock trading volume prediction [32]. However, the increase of model parameters demands more computational resources, limiting their applicability in latency-sensitive scenarios like high-frequency trading (HFT).…”