Online-Tuned Fuzzy Pre-Filtering with an Attention BiLSTM for Misbehavior Detection in Vehicular Named Data...

Area of Science:

• Computer Science
• Network Security
• Artificial Intelligence

Background:

• Vehicular Named Data Networking (VNDN) is vulnerable to position-falsification attacks due to its broadcast-oriented forwarding.
• Existing detection methods use static thresholds or uninterpretable deep models, limiting their effectiveness and trustworthiness.

Purpose of the Study:

• To propose a novel, interpretable, and accurate two-stage detector for position-falsification attacks in VNDN.
• To address the limitations of existing detectors by combining adaptive fuzzy logic with attention-augmented deep learning.

Main Methods:

• A two-stage detection system combining an Adaptive Fuzzy Membership Tuning (AFMT) pre-filter with an attention-augmented bidirectional LSTM (BiLSTM).
• AFMT dynamically updates fuzzy membership functions using gradient descent based on BiLSTM feedback, replacing static thresholds.
• The BiLSTM utilizes the fuzzy suspicion score and provides interpretable attention weights indicating attack onsets.

Main Results:

• The detector achieved high F1-scores (0.955-0.979) on a simulated VNDN benchmark and an F1 of 0.986 on a live testbed.
• It demonstrated competitive performance against strong baselines, achieving the highest ROC-AUC (0.984) and efficient processing (0.44 ms/sample on CPU).
• Transfer learning on the VeReMi v1.0 dataset showed effectiveness across four attack types, highlighting the value of named-data-plane features.

Conclusions:

• The proposed detector effectively identifies position-falsification attacks in VNDN with high accuracy and interpretability.
• The adaptive fuzzy pre-filter and attention-augmented BiLSTM offer a robust and dynamic solution superior to existing methods.
• The findings underscore the importance of named-data-plane features for robust security in vehicular networks.