GATE: Graph Attention-based Ego-Graph Framework for Context-Violating Rogue xApp Detection

Overview

GATE is a research-oriented framework for detecting malicious or faulty xApps in O-RAN environments using context-aware relational learning. The core idea is that an xApp may send syntactically valid control commands through standard interfaces, yet those commands may still be semantically harmful when compared against live network context.

This project focuses on identifying such context-violating control logic by modeling O-RAN behavior as a heterogeneous graph composed of:

• xApps
• UEs
• Cells

The framework uses ego-graph localization and Graph Attention Networks (GAT/GATv2) to detect persistent, targeted, and low-volume malicious behavior that would be missed by signature-based intrusion detection or flat machine learning models.

This work is designed as a containerized, framework-level proof-of-concept, not as a production Near-RT RIC deployment.

---

Problem Statement

Open RAN introduces openness, programmability, and third-party xApps into the Near-RT RIC. While this improves flexibility and innovation, it also expands the attack surface.

A malicious or faulty xApp can:

• issue protocol-compliant E2-like control actions,
• appear valid at the packet or interface level,
• repeatedly target specific UEs or cells,
• and cause persistent degradation in QoS, throughput, SINR, or load balance.

Traditional IDS methods are not sufficient because they often detect malformed packets, known signatures, or coarse traffic anomalies. They do not verify whether an xApp’s action is correct with respect to current network context.

GATE addresses this gap by learning whether xApp behavior is consistent with network state and historical relational patterns.

---

Key Idea

The framework assumes that malicious xApps are not necessarily noisy or obviously disruptive. Instead, they may behave like stealthy logic attackers:

• they may affect only a small subset of UEs,
• they may remain protocol compliant,
• they may avoid triggering threshold-based alarms,
• but they repeatedly make decisions that worsen network conditions under similar contexts.

GATE detects this by learning the relationship between:

• control intent from E2-like action streams,
• network context from SDL-like state,
• interaction metadata from RMR-like messaging behavior.

---

Objectives

• Detect context-violating xApp behavior using radio and system context.
• Model O-RAN interactions as heterogeneous relational graphs.
• Localize inference around each xApp using ego-graph extraction.
• Use graph attention to focus on important xApp–UE interactions.
• Support automated mitigation and reporting as a future extension.
• Maintain reproducibility through time-ordered, rule-labeled synthetic data.

---

System Architecture

The framework is organized into five conceptual modules:

1. Discrete Event Simulation / Data Generation

A synthetic but semantically faithful O-RAN-like environment is generated. It includes:

• xApp actions,
• UE-level radio context,
• cell-level load state,
• xApp telemetry,
• messaging metadata,
• malicious behavior injection.

The generated dataset is:

• time ordered,
• reproducible,
• rule labeled,
• physics consistent,
• suitable for supervised graph learning.

2. Heterogeneous Global Graph Construction

At each timestamp, the system builds a graph where:

• nodes represent xApps, UEs, and Cells
• edges represent:
  • xApp → UE control interactions
  • UE → Cell association relationships

This preserves both behavioral and structural dependencies.

3. Ego-Graph Extraction

Instead of classifying over the full graph directly, each xApp is used as the center of a 2-hop ego graph. This provides:

• xApp-centric localization,
• reduced computation,
• clearer structural representation,
• suitability for low-latency inference.

Each ego graph becomes one supervised sample.

4. Attention-Guided Detection Engine

The detection engine uses GATv2-based graph learning to weigh the importance of neighbors and interactions. It is intended to detect:

• persistent harmful behavior,
• low-volume targeted attacks,
• relational inconsistencies between action and context.

5. Mitigation and Response Policy

A future mitigation module can perform actions such as:

• subscription revocation,
• xApp isolation,
• structured alarm generation,
• escalation for repeat offenders.

At the current stage, mitigation is part of the framework design and not the main implemented contribution.

---

Input Streams

The framework is built around three logical input streams:

E2-like Control Stream

Represents the intent of the xApp.

Examples:

• handover decisions,
• scheduling changes,
• power adjustments,
• control parameter updates.

SDL-like Context Stream

Represents the network state used to evaluate whether the control action is contextually valid.

Examples:

• RSRP,
• SINR,
• throughput,
• QoS score,
• cell load,
• UE count.

RMR-like Interaction Metadata Stream

Represents message-level relational behavior useful for detecting signaling anomalies and interaction patterns.

Examples:

• message type,
• routing key,
• message rate,
• message size,
• execution latency.

---

Dataset Design

The synthetic dataset is designed to reflect realistic O-RAN behavior while remaining controlled and reproducible.

Example Topology

• 15 xApps
• 150 UEs
• 3 Cells
• 6000 timesteps

Example Features

The dataset includes the following classes of features:

Control / Intent Features

• xapp_id
• action_type
• target_ue
• target_cell
• param_value

Radio / Context Features

• rsrp
• sinr
• throughput
• qos_score
• cell_load
• ue_count

Temporal Degradation Features

• sinr_delta
• throughput_delta
• qos_delta

Messaging / RMR-like Features

• msg_type
• routing_key
• rmr_msg_rate
• rmr_msg_size_avg

xApp Telemetry Features

• cpu_usage
• memory_usage
• execution_latency
• prb_usage_dl
• prb_usage_ul

Label

• is_malicious

Labeling Strategy

Labels are assigned through explicit rules, not random assignment. Malicious behavior is injected as:

• repeated targeting of victim UEs,
• intentional degradation of SINR and throughput,
• amplified control intensity or message activity,
• persistent context-violating behavior over time.

This ensures the labels are behaviorally meaningful.

---

Why Graph Attention Networks?

Many traditional approaches are insufficient for this problem:

• Rule-based systems miss stealthy logic attacks.
• Standard ML models treat samples independently and ignore topology.
• Temporal models capture sequences but not multi-entity structure.
• GCNs may average out targeted anomalies through uniform aggregation.

GAT-based models are preferred because they:

• assign different weights to different neighbors,
• preserve relational structure,
• focus on suspicious xApp–UE interactions,
• improve sensitivity to low-volume, high-impact attacks.

This makes GAT especially suitable when one xApp repeatedly harms only a small subset of entities.

---

Model Design

The current framework uses an Attention-Guided Ego-GAT model built using PyTorch Geometric.

High-Level Structure

• multi-head graph attention layer,
• attention refinement layer,
• batch normalization,
• dropout,
• hybrid graph readout,
• fully connected classifier.

Hybrid Readout

The final graph representation combines:

• ego node embedding,
• global max pooling,
• global mean pooling.

This helps capture both:

• highly targeted attacks,
• and more distributed anomalous behavior.

---

Training Setup

Example training configuration from the current implementation:

• Train/Test split: 80/20
• Batch size: 32
• Optimizer: Adam
• Learning rate: 0.005
• Epochs: 20

Reported Baseline Performance

• Accuracy: ~89–90%
• Macro F1: ~0.86–0.87
• Recall: ~0.87

A threshold-tuned configuration was also explored to improve malicious recall.

---