Introduction
As the telecommunications industry transitions toward Open Radio Access Networks (O-RAN), the traditional proprietary "black box" model is being replaced by a disaggregated, software-defined architecture. This shift outsources the "brain" of the network to third-party, AI-driven applications. While this openness fosters innovation and vendor diversity, it introduces a significant security challenge: the vulnerability of these applications to adversarial machine learning. This paper investigates the feasibility of black-box evasion attacks, demonstrating that an adversary—without any knowledge of the underlying AI model—can degrade network performance, bypass defenses, and force cell deactivations in real-time.
The O-RAN Attack Surface
O-RAN introduces the RAN Intelligent Controller (RIC), a centralized platform that hosts third-party applications designed to optimize network functions. These are categorized into:
- xApps: Operating in the Near-Real-Time (Near-RT) RIC, these apps make decisions in sub-second intervals (<1s) for tasks like interference management.
- rApps: Operating in the Non-Real-Time (Non-RT) RIC, these apps handle long-term optimization (>1s), such as power saving and capacity planning.
Because these apps rely on telemetry data (e.g., Key Performance Metrics or spectrograms), they are susceptible to data poisoning and evasion attacks. If an attacker gains access to the data pipeline—potentially through a malicious co-resident app or a side channel—they can inject "noise" into the telemetry data to manipulate the AI's decision-making process.
Methodology: A Four-Stage Black-Box Strategy
The authors propose a tailored attack strategy designed to overcome the unique constraints of the O-RAN environment, particularly the strict timing requirements of the Near-RT RIC.
- Model Cloning (MCA): Since the victim's model architecture is unknown, the attacker observes input-output pairs to train a "surrogate" model. This surrogate serves as a proxy for generating adversarial examples.
- Input-Specific Perturbations: Utilizing techniques like the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD), the authors generate noise tailored to specific data points. While highly effective, these methods are often too computationally intensive for the millisecond requirements of xApps.
- Universal Adversarial Perturbations (UAPs): To address the timing bottleneck, the authors generate a single "master noise" vector. Once computed, this vector can be added to any input with negligible computational overhead, making it ideal for near-real-time evasion.
- Targeted UAPs (TUPs): This refined approach forces the model into a specific, incorrect classification (e.g., reporting a "clear" channel when interference is present) to maximize specific network damage.
Experimental Setup and Datasets
The study evaluated the proposed attacks across three distinct datasets and environments:
- Spectrogram Dataset: 3,000 RGB images representing uplink signals, used for evaluating an Interference Classification (IC) xApp.
- KPM Dataset: 2,910 instances of Key Performance Metrics (SINR, throughput, BLER) collected from a real-world O-RAN testbed.
- City-Scale PRB Dataset: A 40-day dataset of Physical Resource Block (PRB) utilization from a real cellular network, used to evaluate a Power Saving rApp.
Key Insights and Results
The Timing-Effectiveness Tradeoff
The research highlights a critical "timing trap." While iterative attacks like DeepFool are highly effective at lowering model accuracy, they often exceed the execution window of the Near-RT RIC. UAPs solve this problem by providing a pre-computed perturbation that is computationally "free" at runtime.
Attack Effectiveness and Network Impact
In the Interference Classification xApp, the black-box attack dropped model accuracy from 94.8% to 35.8% using only a surrogate model. In over-the-air testing, this caused the network to fail to detect interference, forcing it to use inefficient Modulation and Coding Schemes (MCS). This resulted in significantly higher Block Error Rates (BLER) and a marked decrease in user throughput.
Sabotaging Network Management
The attack on the Power Saving rApp demonstrated even more severe consequences. Using the Keysight RICTest emulator, a targeted attack successfully tricked the system into deactivating capacity cells during peak hours. This led to an immediate and massive drop in downlink throughput for users, effectively sabotaging the network's quality of service.
Resilience of Attacks Against Defenses
The researchers tested their methods against industry-standard defenses, including Adversarial Training and Defensive Distillation. The results showed that these defenses are brittle; the black-box attacks remained successful with only a marginal increase in the perturbation level (noise) required to fool the model.
Conclusion and Future Work
This study proves that O-RAN's openness creates a viable pathway for sophisticated black-box attacks. The ability to clone models and deploy universal perturbations allows attackers to disrupt critical network functions in real-time.
Future research will focus on developing O-RAN-specific defenses, such as runtime behavioral attestation for apps, and extending the threat analysis to Reinforcement Learning (RL) models used for traffic steering and resource scheduling. Additionally, the "stealthiness" of these attacks against network anomaly detectors remains a key area for investigation.
Summary updated