Vehicle intention recognition at signalized intersections based on security-aware inverse reinforcement learningOA

To address the challenge of distinguishing subjective aggressive driving(initiated by drivers)from hazardous behaviors caused by external cyberattacks,this study proposes an innovative intent recognition framework named Intent-Decipher.By integrating the information credibility outputted by an intrusion detection system(IDS)into a security-aware inverse reinforcement learning(SA-IRL)model,the framework infers the reward function behind vehicle behaviors and classifies three key driving intents:normal,aggressive,and malicious.Experiments were conducted on a semi-synthetic dataset containing 20000 trajectories.Results show that Intent-Decipher significantly outperforms baseline methods in classification accuracy,achieving a macro-average F1-score of 0.94.Notably,Intent-Decipher excels at differentiating subjective aggressive driving from attack-induced behaviors:its F1-score for identifying malicious attack-induced(MAI)intent reaches 0.90,an absolute improvement of 0.16 compared with the standard inverse reinforcement learning(IRL)model(which lacks security awareness and only achieves an F1-score of 0.74).