Open Radio Access Network (OpenRAN)

• OpenRAN is a modular network architecture that disaggregates the RAN into RU, DU, and CU, enabling multi-vendor interoperability and flexible deployment.
• It leverages virtualization and network slicing to dynamically allocate resources and support real-time orchestration and adaptive management.
• The open interfaces and AI/ML integration in OpenRAN introduce expanded security challenges, necessitating robust defense measures and advanced anomaly detection.

Open Radio Access Network (OpenRAN, O-RAN) defines a paradigm shift in mobile network architecture, emphasizing disaggregation, openness, and programmability through standardized interfaces, virtualization, and the integration of AI/ML techniques. Evolving from legacy, closed, and vertically integrated RAN deployments, OpenRAN enables multi-vendor interoperability, fine-grained real-time network control, and flexible resource management—particularly critical in next-generation (5G/6G) systems. The architecture decomposes the RAN into modular elements (Radio Unit (RU), Distributed Unit (DU), Central Unit (CU)), each deployable on commodity hardware, with open interfaces (e.g., E2, O1, A1) enabling cross-layer control, observability, and intelligent automation. While these properties accelerate innovation and cost-efficiency, they also introduce a vastly expanded threat surface, new system integration challenges, and the requirement for advanced orchestration primitives.

1. Architectural Disaggregation and Standardization

OpenRAN replaces monolithic, vendor-specific base stations with a modular framework comprising RU, DU, and CU (Motalleb et al., 2019, Masur et al., 2021, Thiruvasagam et al., 2023). The RU executes RF and L1 (PHY-low) tasks, while the DU runs MAC, RLC, and higher PHY. The CU handles RRC, PDCP, and SDAP functions. Both CU and DU are typically implemented as Virtual Network Functions (VNFs) or Cloud-Native Network Functions (CNFs), deployed on commercial off-the-shelf (COTS) hardware—within a cloud environment or edge DC ("O-Cloud") (Thiruvasagam et al., 2023, Azariah et al., 2022).

Disaggregation is operationalized via open, standardized interfaces (notably E2, A1, O1, O2, F1, 7.2x fronthaul), which support:

• Multi-vendor interoperability and plug-and-play component replacement
• Separation of control and user planes, critical for SDN integration
• Modularization of management and orchestration through the Service Management and Orchestration (SMO) layer

Near-RT and Non-RT RAN Intelligent Controllers (RIC) are hosted at different time scales, enabling policy enforcement, closed-loop control, and xApp/rApp-driven innovation (Thiruvasagam et al., 2023, Abdalla et al., 2021). The architecture is sufficiently flexible to support not only Public Land Mobile Networks (PLMNs) but also Non-Public Networks (NPNs), integrated access/backhaul (IAB) (Moro et al., 2023), and future ISAC deployments (Lindenschmitt et al., 21 Sep 2025).

2. Virtualization, Network Slicing, and Adaptive Orchestration

Virtualization underpins OpenRAN’s flexibility, enabling logical network slices to be created atop shared physical infrastructure (Motalleb et al., 2019, Mimran et al., 2022, Zhao et al., 12 Jan 2025). Each slice bundles physical radio resources (PRBs), RUs, and cloud resources (VNFs in DU/CU). Network slicing facilitates multi-service architectures where resources are partitioned according to QoS requirements such that the operation of one service does not impact another. Slices are orchestrated in real-time by open control loops in the RIC, with the ability to adapt resource allocation, power, and placement strategies dynamically (Zhao et al., 12 Jan 2025).

The AdaSlicing architecture (Zhao et al., 12 Jan 2025) encapsulates how adaptive, continual online learning (via Bayesian learning agents and ADMM coordination) can optimize slicing under non-stationary network conditions, implementing soft-isolated virtualization—the sharing of otherwise idle slice resources via adjustable sharing weights and continuous update of allocation strategies—to maximize utilization and performance assurance without sacrificing isolation.

3. AI/ML Integration and Programmable xApps/rApps

OpenRAN explicitly integrates AI/ML to support both long-term and near-real-time decisions (Masur et al., 2021, Thiruvasagam et al., 2023). Non-RT RIC (≥1s timescales) undertakes global optimization, model training, and policy management, while Near-RT RIC (10 ms–1s) deploys xApps for closed-loop control of scheduling, beamforming, load balancing, or anomaly detection (Zhang et al., 2022, Bogucka et al., 13 Mar 2025). xApps implement modular, sharable logic and register via service models (e.g., E2SM-KPM for KPIs, E2SM-RC for RAN Control).

The team learning approach in resource allocation xApps (Zhang et al., 2022) demonstrates the benefits of coordination (sharing intended actions among peer xApps) over independent learning—yielding higher throughput (8% improvement at 6 Mbps load) and much lower packet drop rates (64.8% reduction at 20 m/s mobility) compared to decoupled DQN approaches.

AI/ML-driven applications require robust cross-layer and cross-xApp interoperability, careful conflict mitigation, and mechanisms to ensure explainability and resilience against adversarial attacks. Explainable AI (XAI) technologies (e.g., SHAP, LIME) are gaining adoption for critical functions such as energy management, supporting interpretable AI-driven policies and feature attribution for energy efficiency optimization (Malakalapalli et al., 25 Apr 2025).

4. Security Risks, Attack Surfaces, and Defense Mechanisms

The expansion in functional openness, virtualization, and programmable automation in OpenRAN fundamentally enlarges the attack surface across several domains (Mimran et al., 2022, Liyanage et al., 2022, Chen et al., 2023, Groen et al., 2023, Michaelides et al., 2 Sep 2024):

• Architectural openness exposes new interfaces (O1/O2/E2/7.2x), making them susceptible to man-in-the-middle (MITM), desynchronization, or protocol-level attacks, e.g., rogue O-RU infiltration or falsified KPI injection (Liyanage et al., 2022, Michaelides et al., 2 Sep 2024).
• Virtualization and cloud deployment inherit standard cloud threats: co-residency, VM escape, image poisoning, and supply chain risks (Chen et al., 2023, Liyanage et al., 2022). Network slicing can enable lateral attacks (inter-slice/intra-slice) and needs strong isolation strategies.
• AI/ML threats include poisoning, evasion (adversarial examples), model extraction, and insider attacks via malicious or ill-configured xApps/rApps (Chen et al., 2023, Groen et al., 2023, Liyanage et al., 2022).

Defense approaches span:

• Adoption of cryptographic best practices (TLS, IPSec on open interfaces) (Groen et al., 2023), zero-trust architectures, hardware-backed key storage, and robust authentication.
• Physical-layer device fingerprinting and massive MIMO for device verification (Liyanage et al., 2022).
• AI-driven anomaly detection (autoencoders for KPI deviations), defensive distillation, and adversarial training (Groen et al., 2023, Bogucka et al., 13 Mar 2025).
• Blockchain-enabled distributed trust frameworks for decentralized authentication and mutual attestation between O-RAN components (Liyanage et al., 2022).
• Security-by-design and continuous monitoring of cloud and API resources; automated configuration scanners and sandboxing are recommended (Chen et al., 2023, Liyanage et al., 2022).

Testbed-based work (e.g., (Bogucka et al., 13 Mar 2025)) validates these concepts with practical xApp implementations for jamming and signaling storm detection/mitigation—using sliding window BLER estimation and anomaly score thresholds on timing advance, respectively.

5. Hardware Acceleration, Performance, and Energy Efficiency

Performance constraints at scale necessitate hardware acceleration for Layer 1 processing, especially for PHY-intensive operations (e.g., LDPC/polar coding, massive MIMO, beamforming). The processing load scales with channel bandwidth, antenna count, and inversely with TTI (Kundu et al., 2023): $P \propto \frac{BW \times N_{\text{ant}}}{TTI}$ Accelerators (FPGAs, GPUs, ASICs, SoCs) are incorporated in both lookaside (function-specific) and inline (pipeline-wide) modes, with inline acceleration providing lower latency and reduced CPU bottleneck (Kundu et al., 2023).

On the energy modeling front, OpenRAN system design is constrained by the trade-off between processing centralization (favoring energy efficiency via resource pooling in central DCs) and the increased transmission capacity requirements (and associated energy consumption) for fronthaul eCPRI data when baseband processing is not performed at the edge (Tariq et al., 30 May 2025). LaTeX-modeled expressions are provided for both processing and transmission power, directly informing operator design decisions regarding BBU placement, nodal fanout, and centralized vs. distributed deployment topologies.

6. Experimental Prototyping, Interoperability, and Real-World Applications

A rich ecosystem of open-source projects (srsRAN, OpenAirInterface, O-RAN Software Community) underpins the operability, interoperability, and rapid prototyping of OpenRAN networks (Upadhyaya et al., 2022, Azariah et al., 2022, Thiruvasagam et al., 2023). SDR-based testbeds validate real-world scenarios:

• Integration of E2SM-KPM (metric reporting) and E2SM-RC (control) models for closed-loop mobility load balancing xApps, enabling real-time handover based on comprehensive RAN load metrics (PRB utilization, MAC buffer volume) (Bashar et al., 2 Sep 2025).
• Prototyping IAB over O-RAN by extending standard interfaces to expose IAB-specific telemetry, supporting both centralized and distributed control (Moro et al., 2023).
• Modular ISAC integration using mono-static half-duplex “sniffer” RUs and fronthaul enhancements to support radar functionality with security and minimal hardware changes (Lindenschmitt et al., 21 Sep 2025).

Open-source frameworks and O-RAN Community Labs play a pivotal role in system integration, continuous validation, and cross-vendor deployment, supporting robust academic and industrial testing (Azariah et al., 2022, Upadhyaya et al., 2022).

7. Research Challenges and Future Directions

Despite significant advances, several unsolved issues persist:

• Security: Achieving comprehensive zero-trust architectures with robust cross-domain authentication while mitigating novel attacks on open interfaces and AI/ML components (Abdalla et al., 2021, Liyanage et al., 2022, Groen et al., 2023).
• Latency and Real-Time Control: Addressing deterministic latency guarantees (especially in fronthaul via eCPRI) and sub-millisecond control (e.g., for future URLLC and ISAC) which may demand new RT RICs and hardware co-design (Abdalla et al., 2021, Lindenschmitt et al., 21 Sep 2025).
• Cross-layer Optimization: Tight AI/ML-driven cross-layer orchestration (from physical layer control to cloud-native management) with explainability, privacy, and energy efficiency in mind (Malakalapalli et al., 25 Apr 2025, Zhao et al., 12 Jan 2025).
• Standardization and Interoperability: Evolving APIs, protocol specifications, and security requirements to support more granular function splitting, intelligent testbeds, and sustainable ecosystem growth (Thiruvasagam et al., 2023, Azariah et al., 2022).
• Energy & Resource Efficiency: Optimizing baseband processing placement, hardware acceleration, and dynamic resource slicing to minimize operational costs and carbon footprint (Tariq et al., 30 May 2025, Malakalapalli et al., 25 Apr 2025, Zhao et al., 12 Jan 2025).

Ongoing research is focused on advanced testbeds, digital twin–driven simulation and management, blockchain-enabled control and authentication, and explainable/intelligent orchestration—all essential for resilient, adaptive, and future-proof OpenRAN network deployments.