Comparison of Machine Learning Algorithms Applied to Trust and Position-Based Methods for Malicious Node Detection in Vehicular Ad Hoc Networks

Abstract

Purpose – This study explores the detection of malicious nodes in Vehicular Ad Hoc Networks (VANETs) by means of vehicular trust ratings. It tackles the security and privacy risks of these dynamic networks, acting as the heart of Smart Cities.

Method – Based on the CRISP-DM method, this paper evaluates five algorithms, such as SVC and MLP. The approach specifically tackles data imbalance with Adaptive Synthetic Sampling (ADASYN) and contributes to model transparency with SHapley Additive exPlanations (SHAP), enhancing model interpretability.

Results – Although Support Vector Classifiers (SVC) outperformed conventional accuracy, the improved MLP model trained with hyperparameter tuning for security was found to be better. ADASYN was used to augment the MLP, which achieved a critical recall rate of 1.00 (all nodes were successfully identified as malicious).

Discussion – This is due to the fact that false negatives must be eliminated as a necessity for safety-critical systems. The MLP model’s ability to identify every threat and prevent the general occurrence of a threat from occurring is more valuable than the SVC’s ability to capture threats overall, backed up by elaborate data engineering.

Implication – The results suggest that Explainable AI with SHAP is needed in high-stakes areas such as interconnected roads. Proper accuracy is not enough, and intelligent transportation systems should be transparent and have a high recall for safety and reliability.

Conclusion – The work suggests the importance of a customized tuning for VANET security. Given a recall of 1.00, the MLP model, which is ADASYN-based, protects the network from adversary behavior that is well-documented, showing the class imbalance in the intrusion detection problem.

Recommendation – For intelligent transportation systems designers, system recall should have zero false negatives to make recall the most effective solution to ensure zero false negatives. Explainability frameworks in synthetic sampling can also be integrated into future frameworks to reduce dataset imbalance, ensure that the dataset is unbiased and trustworthy, and make sure that automated decisions made would be dependent and transparent decisions will be performed automatically.