A defence that breaks the moment someone pushes on it.
Modern vehicles run on the Controller Area Network, a protocol built for reliability rather than security, with no native encryption or authentication. Deep-learning intrusion detection is the leading defence, and it inherits two crippling weaknesses.
Imperceptible noise added to network traffic can flip a malicious packet to “benign”. The model is confident, and wrong.
Strip CICIoV2024 of its ~99.75% duplicate records to reflect realistic forensic conditions and a 1D-CNN collapses to F1 below 0.55, while a lightweight Random Forest stays stable.
Can the very technique used to attack these models be repurposed to fix them?
My hypothesis is that adversarial training, folding FGSM- and PGD-crafted examples back into the training set, acts as targeted data augmentation. It manufactures the hard, unique examples that strict de-duplication removes, and finally lets a deep model outperform the Random Forest baseline in realistic, data-scarce conditions.
Five stages, one question.
Strip the dataset back to unique signatures, measure how far the deep model falls, attack it, then fold those attacks back in as training data and re-measure.
Strict data hygiene
De-duplicate CICIoV2024 down to unique CAN-bus attack signatures, recreating the realistic, data-scarce regime where deep models are known to fail.
Baselines
Quantify how far a 1D-CNN degrades against a Random Forest once denied redundant training data.
Adversarial stress tests
Generate FGSM and PGD evasion attacks with the Adversarial Robustness Toolbox.
Adversarial augmentation
A Multi-Strategy Adversarial Training framework that reuses the attacks as hard training examples.
Robustness and superiority
Measure clean accuracy, adversarial accuracy, false-positive rate and compute cost across attack budgets, with stratified K-fold cross-validation keeping rare attack classes represented.
A positive result offers a blueprint for deep learning in any data-sensitive domain where privacy law forces strict de-duplication, such as medical and financial forensics. It also speaks to automotive standards like UN Regulation No. 155.
The work is in progress and nothing here is a published result yet. The code and findings are to be open-sourced and submitted to peer-reviewed venues.