Research On Adversarial Sequence Generation For RNN By Weighted Finite Automaton Abstraction

With the wider application of Recurrent Neural Network(RNN),the existence of adversarial sequences poses a great threat to such safety-critical applications,including autonomous driving behavior prediction models.Adversarial sequence generation for RNN is a method to improve the robustness of RNNs.However,the current existing adversarial sequence generation methods for RNN generally suffer from low efficiency.Therefore,in response to the above problems,around the prediction model of autonomous driving behavior,a WFA-based adversarial sequence generation method for RNN is proposed to help RNN find the adversarial sequence efficiently.First,the construction process of neural networks for autonomous driving behavior prediction is proposed.Before the neural network is constructed,the spatiotemporal data of autonomous driving scenarios are jointly generated by the simulation tools SCENIC and CARLA,and the spatiotemporal data preprocessing method is utilized to clean,integrate,transform and rearrange the data to ensure data quality; During the neural network construction,Integrate Embedding,LSTM,and Temporal Attention to build a neural network model for autonomous driving behavior prediction.Through the case of autonomous cars’ turning at intersections,the validity and rationality of the modeling method are proved.The experimental results show that the modeling method can improve the accuracy of the autonomous driving behavior prediction,and save computing resources.Second,an abstract WFA-based modeling method for autonomous driving behavior prediction is proposed to improve the interpretability of the neural network.By recording the hidden layer of RNN runtime,the state abstraction,the transition abstraction,and the input abstraction from RNN to WFA are realized,and the WFA quintuple is constructed.The innovation of the proposed RNN-WFA abstraction algorithm Fast k-DCP is that it optimizes the clustering process of the k-DCP and relaxes the constraints,which improves the efficiency and scope of application.Through the case of autonomous cars’ turning at intersections,we analyze the operation process of the autonomous driving behavior prediction model(based on the neural network model built above),which proves that the above method is effective and has good versatility,and faster speed.Third,an adversarial sequence generation method based on RNN-WFA abstraction is proposed to improve the robustness of the autonomous driving behavior prediction model.Conducting abstract perturbation to generate an abstract adversarial sequence firstly,and then instantiate it to generate specific adversarial sequences.The case of autonomous cars’ turning at intersections(based on the AVBT dataset)proves the effectiveness of this method,and the comparison with other adversarial attacking algorithms proves the efficiency of the adversarial sequence generation method.The case of autonomous cars’ turning at intersections goes through this article,and we explore the whole life cycle of modeling for autonomous driving behavior prediction,and contribute the autonomous driving dataset AVBT as the benchmark dataset.