Colloquium lecture: 24. November 2026, 10:15 a.m., Towards Robust and Evasion-Resilient Features in ML-Based NIDS for ICT Supply Chain Security

Betreuer/in: Al Sardy

Bild Besprechungsraum 04.137

Cyberattacks which are targeting Information and Communication Technology (ICT) supply chain are becoming
more frequent and severe. To counter these threats, the European Union (EU) has adopted the EU ICT
Supply chain Security Toolbox to identify, assess, and mitigate cybersecurity risks across information and
communication technology supply chains.

vMachine Learning based Network Intrusion Detection System (MLNIDS)can help detect malicious network activities and help in monitoring. However, because of the number of features and redundancy in network traffic, the models become complex and difficult to interpret. Furthermore,
attackers can alter traffic features such as packet padding and timing constraints to bypass detection without
breaking the underlying malicious payload. Traditional feature selection algorithms prioritize detection
accuracy rather than resilience which makes the systems vulnerable to adversarial attacks.
To address these critical challenges, this research proposes a novel framework which combines model
interpretability with robustness against adversarial attacks. The framework begins with a preprocessing step
that removes identifiers such as IP addresses and timestamps to ensure that the model doesn’t learn from the
identifiers and focuses on the network features. The main contribution is the robust feature selection pipeline,
which combines feature importance rankings from three algorithms: SHapley Additive exPlanations (SHAP),
Permutation Importance, and Recursive Feature Elimination (RFE). The mean ranking is then combined with
a Perturbability Score (PS) to estimates how easily an attacker can manipulate a feature.PS is calculated
using five factors: protocol rules, cardinality, correlation between features, attacker control, and granularity.
This approach aims to select features that are both useful for detecting attacks and difficult for attackers to
manipulate. In addition, the supply chain risk score of each feature was defined and was taken into account to
see the impact on robustness.

The proposed framework was evaluated using three datasets representing different network environments:
CIC-IDS2017 for corporate Information Technology (IT) networks, UNSW-NB15 for general IT networks, and
IEC-104 for industrial Operational Technology (OT) environments. Adversarial attacks were simulated by
adding controlled noise to numerical features and probabilistically changing discrete features using an attack
strength parameter. The results showed that models relying only on standard feature importance methods were
highly vulnerable, with Adversarial Success Rate (ASR) reaching 99.90%. In contrast, models using robust
features achieved improved resistance to attacks while maintaining good detection performance on clean data.
The framework achieved a 21.67% robustness gain on CIC-IDS2017, a 9.81% gain on UNSW-NB15and 9.33%
gain on IEC104 datasets.These results were further enhanced by introducing supply chain risk score.
Further analysis showed that selecting features that are difficult to manipulate makes it harder for attackers
to evade detection without affecting their malicious activities. Attempts to change these features may violate
protocol rules or create inconsistencies in network behavior, increasing the cost and difficulty of successful
evasion. Overall, this thesis presents a practical and computationally efficient approach to improving the
robustness of ML-NIDS by combining explainable AI with domain knowledge about feature manipulability.

The proposed framework strengthens resistance to adversarial attacks while preserving detection performance,
offering a promising approach to improving the security of ICT supply chains and supporting European
cybersecurity monitoring requirements.


Raum 04.137, Martensstr. 3, Erlangen

oder

Zoom:
https://fau.zoom-x.de/j/68350702053?pwd=UkF3aXY0QUdjeSsyR0tyRWtLQ0hYUT09

Meeting-ID: 683 5070 2053
Kenncode: 647333