Source-linked AI summary
Integrated LLM-Based Intrusion Detection with Secure Slicing xApp for Securing O-RAN-Enabled Wireless Network Deployments
Joshua Moore, Aly Sabri Abdalla, Prabesh Khanal, Vuk Marojevic
TL;DR
O-RAN’s modularity and slicing create security challenges, including vulnerabilities in resource-allocation policies. This paper integrates LLM-based detection with secure slicing in an OAIC near-RT RIC deployment; instruction tuning raises Gemma 2 task accuracy from 87% to 99%.
Problem
O-RAN slicing can be vulnerable to attacks exploiting absent robust authentication in resource-allocation policies, disrupting fair resource distribution for connected devices.
Source: §IV · p. 5Method
The framework combines KPI collection, LLM-based UE traffic assessment, and secure slicing that moves an anomalous UE to a zero-resource slice.
Source: §I · p. 2, A. LLM-based Intrusion Detection and… · p. 3Results
Gemma 2 accuracy increased from 87% with demonstration-based prompting to 99% after fine-tuning, and average detection-and-response time was 239.66 ms.
Source: §IV · p. 5Takeaways & Limitations
The OAIC deployment demonstrates an integrated approach intended to protect legitimate O-RAN users by detecting intruders and isolating them through secure slicing.
Source: §V · p. 6
Abstract
from arXiv · showhide
The Open Radio Access Network (O-RAN) architecture is reshaping telecommunications by promoting openness, flexibility, and intelligent closed-loop optimization. By decoupling hardware and software and enabling multi-vendor deployments, O-RAN reduces costs, enhances performance, and allows rapid adaptation to new technologies. A key innovation is intelligent network slicing, which partitions networks into isolated slices tailored for specific use cases or quality of service requirements. The RAN Intelligent Controller further optimizes resource allocation, ensuring efficient utilization and improved service quality for user equipment (UEs). However, the modular and dynamic nature of O-RAN expands the threat surface, necessitating advanced security measures to maintain network integrity, confidentiality, and availability. Intrusion detection systems have become essential for identifying and mitigating attacks. This research explores using large language models (LLMs) to generate security recommendations based on the temporal traffic patterns of connected UEs. The paper introduces an LLM-driven intrusion detection framework and demonstrates its efficacy through experimental deployments, comparing non fine-tuned and fine-tuned models for task-specific accuracy.
I. INTRODUCTION
O-RAN combines flexible, intelligent network management and tailored slices, but its open, modular design creates security risks. This work proposes an LLM-based intrusion-detection framework integrated with secure slicing and evaluates it through an OAIC deployment.
I. INTRODUCTION
- O-RAN disaggregates CU, DU, and RU functions and uses near-RT and non-RT RICs for AI-driven optimization and management.The paper notes that decoupling hardware and software supports multi-vendor deployments and rapid adaptation.Source: §I · p. 1
- Network slicing partitions networks into isolated slices tailored to applications or QoS requirements, while RICs optimize their resource allocation.This can improve resource utilization and service quality for connected UEs.Source: §I · p. 1
- O-RAN’s open interfaces, over-the-air links, and intelligent controllers create attack opportunities that threaten network integrity and availability.The paper identifies intrusion detection as a critical component for identifying and mitigating attacks.Source: §I · p. 1
- The study explores LLM-generated security recommendations based on connected UEs’ temporal traffic patterns to support secure slicing on shared O-RAN infrastructure.Its stated goal is to support heterogeneous traffic types.Source: §I · p. 1
- The proposed near-RT RIC framework combines KPIMON, LLM-based intrusion detection, and secure-slicing xApps to collect data, detect malicious activity, and isolate intruders.The paper describes an experimental OAIC deployment intended to safeguard legitimate users in the presence of intruders.Source: §I · p. 2
II. BACKGROUND AND PRIOR WORK
Prior work covers conventional intrusion-detection methods and LLM-based analysis of 5G traffic, but the cited studies do not integrate the proposed approach into O-RAN. This work targets that deployment context.
II. BACKGROUND AND PRIOR WORK
- O-RAN intrusion-detection systems can combine signature, anomaly, and behavior-based analysis, then respond by isolating components, alerting operators, or adapting configurations.The described data sources include traffic patterns, signaling messages, and system logs from the RU, DU, and CU.Source: §II · p. 2
- Earlier LLM studies explored 5G feature extraction and anomaly detection, but the cited feature-extraction work did not integrate its approach into a 5G network and used closed-source models.The cited federated approach traded slightly lower accuracy than centralized approaches for enhanced privacy and scalability.Source: §II · p. 2
- A cited Llama 2 study reported DDoS detection on CIC-IDS2017 with performance comparable to LSTM, CNN, and DNN models, and suitability for real-time monitoring.The passage does not provide a numeric accuracy value for this comparison.Source: §II · p. 2
- This paper distinguishes its work from prior anomaly-detection or feature-extraction studies by proposing an LLM-driven intrusion-detection framework specifically for O-RAN.Source: §II · p. 2
A. LLM-based Intrusion Detection and Mitigation
LLM-ID is designed for real-time O-RAN anomaly detection and response, using a near-RT RIC xApp pipeline that connects KPI collection and LLM analysis to secure slicing.
A. LLM-based Intrusion Detection and Mitigation
- LLM-ID uses LLM reasoning for real-time anomaly detection and response, with the paper positioning the framework for O-RAN deployment.The authors state that fine-tuning can improve task-specific security-recommendation accuracy for securely sliced RANs.Source: A. LLM-based Intrusion Detection and… · pp. 2–3
- KPIMON collects RAN information, LLM-ID analyzes metrics such as TX packets and UE counts, and SSxApp responds to alerts with security measures.RMR carries messages between xApps, and a database stores KPIMON information for use by other xApps.Source: A. LLM-based Intrusion Detection and… · p. 3
- The framework uses E2-node KPM reports stored by KPIMON; LLM-ID extracts relevant KPIs and assesses intrusion with an instruction-tuned prompt.The passage describes the KPIMON-to-database-to-LLM-ID flow.Source: A. LLM-based Intrusion Detection and… · p. 3
- When the model labels a UE malicious, LLM-ID sends an RMR request to SSxApp to unbind it from its slice and bind it to a zero-resource slice.For a legitimate UE, the model proceeds to another prompt.Source: A. LLM-based Intrusion Detection and… · p. 3
B. Experimental Deployment and Fine-Tuning
The experimental deployment extends OAIC’s near-RT RIC with monitoring and secure-slicing capabilities in a 5G testbed. The setup uses one base station and three eMBB UEs sharing a slice.
B. Experimental Deployment and Fine-Tuning
- The OAIC testbed extends an openly accessible platform for monitoring, analyzing, and securing network slices through xApps in its near-RT RIC.The RIC is based on the O-RAN Software Community’s RIC.Source: B. Experimental Deployment and… · p. 3
- The testbed includes a 3GPP-compliant 5G RAN and core using srsRAN and Open5GS, plus an O-RAN control network with a near-RT RIC and E2 KPM and control-action flows.It supports reconfigurable slicing and scheduling through modular APIs.Source: B. Experimental Deployment and… · p. 3
- The wireless deployment uses one base station and three B210-USRP UEs in one eMBB slice, with 20 MHz bandwidth and 100 resource blocks divided equally among users.The stated eMBB demand is 10 Mbps; the modified RIC includes KPIMON and a custom SSxApp for monitoring and resource isolation.Source: B. Experimental Deployment and… · p. 4
1) Data Collection:
The collection process gathers RAN performance and connection metrics, then preprocesses them to create data for fine-tuning.
1) Data Collection:
- KPIs include packet transmission rates, resource-allocation metrics, and UE connection status, collected through the near-RT RIC.Scripts automate logging; Python tools support analysis and visualization.Source: 1) Data Collection · p. 4
- Legitimate UEs are moved relative to the base station to vary modulation and coding schemes in the collected data.The attacker remains at a fixed distance to keep signal propagation consistent.Source: 1) Data Collection · p. 4
- Python preprocessing removes obvious errors and extracts key parameters for LLM fine-tuning.Source: 1) Data Collection · p. 4
2) Fine-Tuning the LLM:
Gemma 2 is fine-tuned offline on labeled normal and malicious traffic using O-RAN KPIs to support task-specific intrusion detection.
2) Fine-Tuning the LLM:
- Gemma 2 is fine-tuned on labeled normal and malicious traffic, with randomized device and packet-count variables intended to broaden traffic patterns and reduce overfitting.Fine-tuning is performed offline on a GPU-accelerated server.Source: 2) Fine-Tuning the LLM · p. 4
- The tuning uses standardized O-RAN KPIs, including byte counts, resource blocks, packet counts, errors, and active UE counts.Source: 2) Fine-Tuning the LLM · p. 4
- The authors state that aligning training objectives with OTA-tested KPIs supports intrusion identification and security recommendations relevant to real-world deployments.Source: 2) Fine-Tuning the LLM · p. 4
IV. EXPERIMENTAL DEPLOYMENT AND RESULTS
The experiments compare prompted LLMs with task-fine-tuned Gemma 2 and test integrated detection and secure slicing on OAIC. Fine-tuning improves Gemma 2 accuracy, while the deployment detects and isolates malicious UEs to protect legitimate users.
IV. EXPERIMENTAL DEPLOYMENT AND RESULTS
- 99% accuracy after fine-tuning, versus 87% for Gemma 2 with demonstration-based prompting, on anomalous-traffic detection.Fig. 4 compares accuracy with few-shot prompting and fine-tuning.Source: Figure 4 · p. 4, §IV · p. 5
- The experiment simulates a DoS attack in which a malicious UE floods requests and degrades legitimate users' service by exploiting absent robust authentication in slicing resource-allocation policies.Source: §IV · pp. 4–5
- Chat GPT 4 achieved 100% accuracy, followed by Chat GPT 3.5 at 95%, Gemma 2 at 87%, and Llama at 70% with few-shot prompting.Fig. 4a compares model accuracy for anomalous-traffic detection.Source: Figure 4 · p. 4, §IV · p. 5
- The LLM identifies abnormality in a UE’s Tx packets and throughput, classifies it as an intruder, and the secure-slicing xApp moves it to a slice with no resources to mitigate further damage.Fig. 5 shows three UEs sharing a slice with maximum downlink throughput of 10 Mbps; around 170 seconds, one UE becomes greedy and takes resources from the others.Source: §IV · p. 5
- 239.66 ms is the reported average detection-and-response time, measured from receipt of KPM reports through mitigation.KPM reports arrive at intervals between 1 ms and 1 second.Source: §IV · p. 5
- 175 s versus 280 s: the LLM-ID identifies anomalous UE1 promptly, while static-threshold UE2 is not marked malicious until 280 s.Anomalous behavior begins around 175 s; the static method's delay is attributed to false positives and collecting multiple KPM reports.Source: §IV · pp. 5–6, Figure 6 · p. 5
V. CONCLUSIONS
The paper presents an OAIC-tested O-RAN framework combining LLM-based intrusion detection, KPM collection, and secure slicing to detect anomalies and isolate intruders.
V. CONCLUSIONS
- OAIC testbed experiments demonstrate improved task-specific anomalous-traffic accuracy for fine-tuned open-source models.The conclusion presents this as evidence of the potential of LLM-driven frameworks for O-RAN security and resilience.Source: §V · p. 6