Deep Fusion Intelligence: Enhancing 5G Security Against Over-the-Air Attacks

Abstract

With the increasing deployment of 5G networks, the vulnerability to malicious interference, such as jamming attacks, has become a significant concern. Detecting such attacks is crucial to ensuring the reliability and security of 5G communication systems Specifically in CAVs. This paper proposes a robust jamming detection system addressing challenges posed by impairments, such as Carrier Frequency Offset (CFO) and channel effects. To improve overall detection performance, the proposed approach leverages deep ensemble learning techniques by fusing different features with different sensitivities from the RF domain and Physical layer namely, Primary Synchronization Signal (PSS) and Secondary Synchronization Signal (SSS) cross-correlations in the time and the frequency domain, the energy of the null subcarriers, and the PBCH Error Vector Magnitude (EVM). The ensemble module is optimized for the aggregation method and different learning parameters. Furthermore, to mitigate the false positive and false negative, a systematic approach, termed Temporal Epistemic Decision Aggregator (TEDA) is introduced, which elegantly navigates the time-accuracy tradeoff by seamlessly integrating temporal decisions, thereby enhancing decision reliability. The presented approach is also capable of detecting inter-cell/inter-sector interference, thereby enhancing situational awareness on 5G air interface and RF domain security. Results show that the presented approach achieves the Area Under Curve (AUC) of 0.98, outperforming other compared methods by at least 0.06 (a 6% improvement). The true positive and negative rates are reported as 93.5% and 91.9%, respectively, showcasing strong performance for scenarios with CFO and channel impairments and outperforming the other compared methods by at least 12%. An optimization problem is formulated and solved based on the level of uncertainty observed in the experimental set-up and the optimum TEDA configuration is derived for the target false-alarm and miss-detection probability. Ultimately, the performance of the entire architecture is confirmed through analysis of real 5G signals acquired from a practical testbed, showing strong agreement with the simulation results.