Open-RAN (O-RAN) Overview

Updated 30 November 2025

Open-RAN is a disaggregated, cloud-native radio access network architecture that decouples hardware and software with open, vendor-neutral interfaces.
It employs near and non-real-time RAN Intelligent Controllers to integrate AI/ML for dynamic network orchestration and closed-loop control.
The framework enhances resource management, network slicing, and security, validated through both production and experimental deployments.

Open Radio Access Network (O-RAN) constitutes a comprehensive redesign of the cellular RAN, defining an open, disaggregated, cloud-native, and intelligence-augmented paradigm for 5G, 6G, and beyond. By decoupling hardware from software and opening vendor-agnostic interfaces between the main RAN functions, O-RAN enables a multi-vendor ecosystem, advanced orchestration of network services, and tight integration of machine intelligence. Its architectural, operational, and security models are under active standardization by the O-RAN Alliance and have been validated in both production and experimental deployments (Masur et al., 2021, Thiruvasagam et al., 2023, Polese et al., 2022, Alam et al., 6 May 2024).

1. Disaggregated Architecture and Key Components

O-RAN disaggregates the monolithic base station (gNB/eNB) into three standardized, interoperable elements running as virtual network functions (VNFs) or containerized network functions (CNFs) over commercial-off-the-shelf (COTS) hardware: the Radio Unit (O-RU), Distributed Unit (O-DU), and Centralized Unit (O-CU) (Masur et al., 2021, Gopal et al., 15 Nov 2024, Alam et al., 6 May 2024).

O-RU: Performs analog/RF and Low-PHY (FFT/IFFT, A/D-D/A) functions. Sits at the antenna site, terminating the Open Fronthaul interface.
O-DU: Hosts High-PHY, MAC, and RLC, and is typically edge-deployed. Real-time scheduling and HARQ are processed here, with fronthaul (eCPRI) links to multiple RUs.
O-CU: Implements RRC, PDCP, and SDAP for session management and higher-layer processing. Split into CU-CP/UP for control/user plane separation; typically data center deployed.

Above these, RAN Intelligent Controllers (RICs) are introduced:

Near-Real-Time RIC (Near-RT RIC): Operates in 10 ms–1 s loops, hosts xApps for RRM, closed-loop scheduling, and SLA enforcement. Interfaces with O-DU/CU over the E2 interface.
Non-Real-Time RIC (Non-RT RIC): Runs with >1 s periodicity, hosts rApps for policy, ML model training, slice lifecycle, and analytics. Communicates with near-RT RIC (A1 interface) and SMO (O1/O2).

The orchestration, lifecycle management, and northbound network management functionalities are provided by the Service Management and Orchestration (SMO) framework, which exposes the O1 (FCAPS management), O2 (cloud control), and A1 (policy/ML) interfaces (Alam et al., 6 May 2024).

2. Open Interfaces and Control Planes

O-RAN specifies open, vendor-neutral interfaces that enable plug-and-play interoperability, elastic scaling, and network programmability (Polese et al., 2022, Chen et al., 2023, Gopal et al., 15 Nov 2024):

Interface	Connectivity	Primary Purpose
Open Fronthaul	O-DU ↔ O-RU	eCPRI transport (IQ/U-plane, C-plane, S-plane, M)
E2	Near-RT RIC ↔ O-DU/O-CU	Real-time control/telemetry, xApp actions
A1	Non-RT RIC ↔ Near-RT RIC	Policy, ML model updates, long-term intent
O1	SMO ↔ RAN nodes (DU/CU/RICs)	FCAPS mgmt, configuration, performance mgmt
O2	SMO ↔ O-Cloud	Resource provisioning, VNF/CNF lifecycle
F1	O-CU ↔ O-DU	User and control-plane split per 3GPP NG-RAN

This layered and open interface fabric underpins the O-RAN control-plane abstraction: microservices (xApps/rApps) subscribe to KPIs, maintain control feedback loops, and issue commands to data-plane VNFs.

3. Virtualization, Orchestration, and Slicing

O-RAN is fully cloud-native by design: RAN components are deployed as VNFs or CNFs under a virtualized infrastructure (O-Cloud), managed by the SMO using the O2 interface. This enables elastic resource allocation, rapid updates via CI/CD workflows, and decoupling of hardware/software innovation cycles (Masur et al., 2021, Thiruvasagam et al., 2023, Gopal et al., 15 Nov 2024).

Network slicing is realized as follows (Alam et al., 6 May 2024, Gopal et al., 15 Nov 2024):

RAN slice subnets are instantiated by the RAN-NSSMF (slice subnet management function) in the SMO, assigning resources per slice (S-NSSAI) using configuration via O1 and runtime adjustments via E2.
At the O-DU MAC layer, slice-specific PRB shares $P_i = \alpha_i P_\text{total}$ are enforced, with $\sum_i \alpha_i = 1$ . xApps use control primitives to dynamically reshape these shares in response to KPI feedback.
End-to-end orchestration includes coordination with transport and core slices through NSMF/NFMF, with all lifecycle stages—creation, modification, activation, deactivation, termination—instrumented over O1/E2/A1.

Cloud-native orchestration further allows multi-tenancy, fine-grained isolation, and the realization of enterprise/private RAN deployments.

4. Embedded Intelligence and AI/ML Workflows

O-RAN integrates machine intelligence as a first-class architectural element (Masur et al., 2021, Polese et al., 2022, Alam et al., 6 May 2024):

Non-RT RIC: Aggregates historical KPIs, conducts offline ML model training (including supervised, RL, self-supervised, federated learning), and distributes models/policies to the near-RT RIC via A1.
Near-RT RIC: Hosts inference engines (xApps) executing control policies (e.g., beam management, traffic steering, slice scaling) in closed loop (10 ms–1 s). Inference and policy actions are enforced over E2.
AI/ML workflows are built atop a data pipeline collecting telemetry via O1/E2, pre-processing data (normalization, autoencoders), training models (see e.g., reinforcement-learning reward $R_t = \alpha T_t - \beta L_t$ ), benchmarking for generalization and safety, deploying via containers or ONNX artifacts, and performing continuous monitoring for drift and SLA violations.

Experimentation on testbeds (e.g., AERPAW (Moore et al., 6 Nov 2024)) confirms the architecture: FlexRIC/xApps running in Docker on edge servers manage real-time closed-loop scheduling for UAVs and terrestrial UEs, with experimental metrics including throughput, latency, PER, and seamless integration with digital twins (via GANs for data augmentation). Orchestrators like OREO optimize xApp deployment and resource allocation, using multi-layer graph models and Lagrangian heuristics, achieving up to 30% resource reduction compared to monolithic approaches (Mungari et al., 28 May 2024).

5. Security Architecture, Threat Models, and Countermeasures

O-RAN’s open, disaggregated, and cloudified design increases the systemic attack surface and introduces new security and privacy risks (Chen et al., 2023, Groen et al., 2023, Abdalla et al., 2023, Mimran et al., 2022, Groen et al., 23 Apr 2024):

Attack surfaces: Unauthorized access and session hijacking on A1/E2/O1; eavesdropping and tampering on fronthaul (eCPRI); adversarial attacks on ML models; supply chain and image manipulation; hypervisor/VM/namespace escape; misconfiguration; API surface exploits; inter-slice leakage in slicing.
Specific threats:
- Physical hijack or eavesdropping on fronthaul (Open FH, S-plane/SyncE/PTP).
- xApp/rApp ML model poisoning, adversarial example generation, or policy manipulation.
- Container breakout and lateral movement on O-Cloud nodes.
Countermeasures:
- Universal mutual authentication and encryption for open interfaces (TLS 1.3/IPsec/MACsec/X.509/IEEE 802.1X), secure boot, hardware root-of-trust (TPM/SGX).
- Container isolation and runtime integrity monitoring (e.g., SELinux, gVisor, Kata).
- Robust ML: adversarial training, anomaly detection (autoencoders), explainability, continuous validation.
- Supply-chain security: signed images, SBOMs, strict image management, zero-trust privilege principles.
- Network slicing: per-slice keying, resource isolation, and cross-slice access controls.
- End-to-end security audits, plug-fests, and continuous threat modeling integrated into CI/CD (Groen et al., 23 Apr 2024, Abdalla et al., 2023, Polese et al., 2022, Chen et al., 2023).

Experimental results show manageable overheads for strong encryption (E2: +22 μs, 1.37 Gbps with AES-GCM; fronthaul: +39–218 μs, CPU bottlenecks if not hardware-accelerated); autoencoder defenses cut adversarial xApp deviations to <10% (Groen et al., 23 Apr 2024, Groen et al., 2023).

6. Performance, Resource Management, and Energy Efficiency

O-RAN deployments are challenged by real-time constraints (e.g., ≤1 ms for URLLC), virtualization overheads, resource scheduling under multi-vendor heterogeneity, and energy efficiency (Masur et al., 2021, Abughazzah et al., 10 Mar 2025, Tariq et al., 30 May 2025):

Performance modeling: End-to-end latency is modeled as $L_\text{total} = L_\text{fronthaul} + L_\text{processing} + L_\text{backhaul}$ ; resource allocation under security and computational constraints is formalized as a multi-objective optimization problem, balancing latency and security (Abughazzah et al., 10 Mar 2025).
Resource management: Algorithms dynamically associate UEs to RUs and select encryption schemes subject to battery budgets, processing capacity, and security requirements, using convex-iterative hybrid solvers with near-optimal latency and security performance.
Energy models: Transaction-based frameworks decompose power consumption into processing and transport, quantifying trade-offs between baseband processing centralization (lower $P_\text{pr}$ due to sharing in DC) and transport overhead (higher $P_\text{tr}$ for deep centralization) (Tariq et al., 30 May 2025).
Programmability: Exposed scheduler parameters (e.g., PF exponents β in TailO-RAN (Longhi et al., 16 Aug 2025)) controlled by xApps enable tailored SLA compliance (e.g., per-UE throughput/F1 score for IIoT cameras) with tight near-RT control loops (<100 ms from requirement to effect).

7. Standardization, Open Source, and Research Directions

O-RAN standardization is conducted by the O-RAN Alliance (WG1–WG11), in liaison with 3GPP, TIP, and experimental testbeds (Thiruvasagam et al., 2023, Alam et al., 6 May 2024):

O-RAN WG	Focus Area
WG1	Architecture & use cases
WG2	Non-RT RIC & A1 interface
WG3	Near-RT RIC & E2 interface
WG4	Open Fronthaul (CUS/M/S-Plane specs)
WG5	F1/E1/Nx APIs, O1 (mgmt), MH/BH profiles
WG6	Cloud/Orchestration (O2, O-Cloud design)
WG7	White-box hardware (RU/DU/indoor/outdoor)
WG8	Base station stack & API reference
WG9	Transport (X-haul, DWDM, slicing)
WG10	OAM (O1 mgmt models, function)
WG11	Security (threat modeling, interface sec)

Open source ecosystems (O-RAN SC, ONAP, OSM, OpenAirInterface, srsRAN, OpenRAN Gym, Colosseum, Arena, AERPAW) foster experimental validation, rapid prototyping, and multi-vendor plug-fests (Upadhyaya et al., 2022, Moore et al., 6 Nov 2024, Mungari et al., 28 May 2024).

Research challenges: Multi-RIC coordination, real-time ML at PHY, AI explainability and robustness, scalable orchestration, end-to-end cross-domain slicing, low-latency cryptographic primitives, zero-touch trust anchors, digital twin–based testing, blockchain for provenance, and non-terrestrial/satellite RAN integration are identified as future directions (Abdalla et al., 2021, Gopal et al., 15 Nov 2024, Michaelides et al., 2 Sep 2024).

In summary, O-RAN establishes a programmable, disaggregated, AI-native, and open RAN paradigm, coupling multi-vendor flexibility, advanced slicing, and security-by-design principles. Realization at scale demands persistent effort in standardization, open-source validation, secure AI/ML integration, and end-to-end orchestration for future 5G/6G networks (Masur et al., 2021, Alam et al., 6 May 2024, Gopal et al., 15 Nov 2024, Abughazzah et al., 10 Mar 2025, Longhi et al., 16 Aug 2025).