Comparative Analysis of Recurrent Neural Network Architectures for Real-Time VANET Misbehavior Detection: Performance vs. Efficiency Trade-offs

Abstract

It is still challenging to find inappropriate behavior in Vehicular Ad-hoc Networks (VANETs) since it requires conducting such actions promptly and properly, particularly in safety scenarios that occur on the fly. This paper provides a comprehensive list of three Recurrent Neural Network (RNN) systems, LSTM, GRU, and Bidirectional LSTM, which were tested on 20 forms of attacks. We review the trade-off between detection accuracy and computational efficiency using the VeReMi Extension dataset that comes in sample sizes of 500k to 2M. The most accurate (97.02% with 500K samples) is Bidirectional LSTM which is 15.8% slower to train than the standard LSTM. GRU took 282 minutes to train on 2 million samples and achieved an accuracy of 89.03%. The latency of inference of all of the models is under 41 ms, indicating an ability to be used immediately. Bi-LSTM reduces the missed-attack rate by 5.04%-2.32%, that is, it is a good substitute in looking for places where safety is a key consideration. These outcomes indicate the way to choose an appropriate model architecture under various circumstances.