Contrastive Learning for Distortion Tolerable Network Slice Prediction in Open RAN

Open Radio Access Network (Open RAN) has revolutionized future communications by introducing open interfaces and intelligent network management. Network slicing enables the creation of multiple virtual networks on a single physical infrastructure, providing tailored services for performance, security, and latency. Efficient RAN slice resource allocation requires accurate prediction of the slice loads from the collected reports. However, open interfaces brought by Open RAN have also caused new security challenges. Malicious attackers could modify the data between E2 nodes with Near Real-Time RIC, hence mislead the model for a poor performance. To prevent this attack, we hereby proposed a novel contrastive learning design, which uses data augmentation to grant the model the vulnerability of feature distortion. The contrastive learning model could learn the correlation of original data with distorted data. Meanwhile, the proposed contrastive learning has a greater generalization ability compared to conventional supervised learning, which is suitable for dynamic environments and could adapt to various noise levels. The proposed contrastive learning includes supervised and unsupervised contrastive learning (SCL and UCL). The proposed SCL could achieve 87.1% out-of-distribution network slice classification accuracy, the proposed UCL could achieve 86.6%, while the conventional MLP is 82.6%. Meanwhile, the proposed method only requires 8.4% of computation during training compared to that of conventional MLP.