Hunting Adversaries in Federated IDS: Does Local Adversarial Training Actually Make the Global Detector Harder to Evade?

The objective of the project is to enhance the resilience of Federated Intrusion Detection Systems (FIDS) against adversarial attacks. Local Adversarial Training (LAT) will be used so that every client can develop and train on weak adversarial traffic in the local environment to enhance the robustness of the global model against potential poisoning and evasion attempts. The evaluation of this defense will include realistic federated contexts, including non-IID data and models undermined by a Byzantine participant. The project will also incorporate two dimensions into the evaluation of the system by combining machine learning robustness metrics and operational cybersecurity metrics—such as MITRE ATT&CK mappings, model drift detection, and time-to-detection—to better indicate how the AI robustness will operate in a real-world SOC context. The findings will result in a concrete basis for the development of privacy-preserving, trustworthy intrusion detection systems, while also contributing to the state-of-the-art in trust architectures for AI and resilience of Canada’s digital infrastructure.

Faculty Supervisor:

Chadi Assi

Student:

Partner:

Ecole supérieure en Sciences et Technologies de l'Informatique et du numérique (ESTIN)

Discipline:

Computer science

Sector:

Education

University:

Concordia University

Program: