Detecting and Mitigating Adversarial Machine Learning in Open RAN Using Digital Twin

This letter proposes a Digital Twin (DT)-based solution to detect and mitigate Adversarial Machine Learning (AML) attacks in the Open Radio Access Network (Open-RAN). Since ML models in xApps and rApps within the RAN Intelligent Controller (RIC) are vulnerable to AML, we use a sliding-window technique to compare outputs between the physical system (PS) and DT. Significant discrepancies trigger the replacement of the model with a secure DT-trained version. Our approach also localizes AML attacks within the network. A case study on a DL-based xApp validates its effectiveness against AML attacks such as the fast gradient sign method (FGSM) and query-based attacks.