Technical Summary: AI-driven Intrusion Detection for UAV in Smart Urban Ecosystems

1. Problem Statement

The integration of Unmanned Aerial Vehicles (UAVs) into Smart Urban Ecosystems offers transformative potential for applications such as traffic monitoring, disaster response, and delivery services. However, this proliferation introduces a dual-layer security crisis that existing literature often addresses in isolation:

1. Cyber-Attacks: UAVs, as cyber-physical systems, are vulnerable to attacks targeting their communication infrastructure (e.g., jamming, spoofing, message injection, and denial-of-service). These exploits can lead to loss of control, data theft, or mission failure.
2. Physical Intrusions: Conversely, unauthorized UAVs themselves pose threats to public safety and privacy. They can be used for espionage, disrupting public events, interfering with air traffic (including emerging eVTOLs), or releasing hazardous materials.

Current survey literature typically focuses on either cyber Intrusion Detection Systems (IDS) or counter-UAV physical sensing, lacking a unified framework that addresses both threats simultaneously within the resource-constrained, dynamic environment of a smart city. Furthermore, there is a scarcity of UAV-specific datasets, with many studies relying on generic IT network data that lacks ecological validity for UAV command-and-control (C2) links and telemetry.

2. Methodology and Framework

This paper provides a structured, evidence-based synthesis of the state-of-the-art in AI-driven IDS for UAVs. The methodology involves:

• Literature Synthesis: A comprehensive review of literature from the last decade (with emphasis on 2019–2025) to categorize threats and solutions.
• Unified Taxonomy Construction: The authors propose a unified IDS architecture that integrates cyber and physical detection. This architecture is organized across four deployment tiers:
  • Onboard: Lightweight detection for flight safety and immediate link integrity (e.g., IMU/GPS consistency checks).
  • Cooperative/Swarm: Peer-to-peer validation to reduce false alarms and detect localized spoofing.
  • Edge/Fog: Intermediate processing for multi-sensor fusion (RF, Vision, Acoustic) and pre-filtering before cloud transmission.
  • Ground Control Station (GCS)/Cloud: Centralized analytics, fleet-wide policy enforcement, forensic analysis, and model retraining.
• AI Technique Analysis: The survey categorizes AI approaches into:
  • Supervised Learning: Using labeled datasets (e.g., SVM, Decision Trees, Deep Learning) to classify known attacks (spoofing, hijacking) or detect unauthorized UAVs via computer vision (e.g., YOLO, CNNs).
  • Unsupervised Learning: Using unlabeled data (e.g., Autoencoders, One-Class SVM) to detect anomalies and unknown threats without prior labeling.
  • Reinforcement Learning (RL): Employing agents to learn optimal detection and mitigation policies in dynamic environments, balancing detection accuracy with latency and energy constraints.
• Data Fusion Strategy: The paper advocates for multi-modal fusion (combining network traffic, RF signatures, visual data, and acoustic sensors) to mitigate environmental challenges (clutter, weather) and adversarial evasion.

3. Key Contributions

The paper makes the following specific contributions to the field:

1. Unified Scope: It integrates the perspectives of cyber-attacks on UAV communication links and physical intrusions by unauthorized UAVs into a single IDS view, addressing a gap in prior surveys that treated these domains separately.
2. Categorization of Threats: It classifies intrusions into two primary classes: cyber-attacks targeting infrastructure and unwanted/unauthorized physical intrusions, detailing specific attack vectors (e.g., GPS spoofing, jamming, privacy invasion).
3. Unified IDS Taxonomy: It presents a detailed taxonomy aligned with sensing, calibration, fusion, decision, and response mechanisms across onboard, cooperative, edge, and ground control tiers.
4. Dataset Consolidation: It consolidates a list of publicly available datasets suitable for both network traffic (e.g., CIC-IDS2018, NSL-KDD) and vision modalities (e.g., Anti-UAV, DUT Anti-UAV), explicitly noting their applicability and limitations for UAV-specific tasks.
5. Research Directions: It identifies ten critical open research directions: Scalability, Robustness, Explainability, Data Scarcity, Automation, Hybrid Detection, Large Language Models (LLMs), Multimodal approaches, Federated Learning, and Privacy Preservation.

4. Results and Findings

The survey synthesizes findings from numerous studies to highlight current capabilities and limitations:

• Performance of AI Models:
  • Supervised Methods: Deep learning and computer vision techniques (e.g., YOLOv4, CNNs) show high accuracy in detecting and tracking UAVs in visual data. For cyber threats, algorithms like Random Forest and XGBoost have demonstrated high accuracy (e.g., ~90% for hijacking detection, ~89.7% for GPS spoofing) on specific datasets.
  • Unsupervised Methods: Approaches like Autoencoders and One-Class SVMs are effective for detecting unknown anomalies and eavesdropping, though they often lack specific numerical benchmarks in current literature.
  • Reinforcement Learning: RL-based IDS (e.g., DQN, DRL-BWO) show promise in dynamic environments, achieving high F1-scores (up to 0.988) in simulation, though often relying on generic cybersecurity datasets rather than UAV-native traffic.
• Deployment Trade-offs:
  • Latency: Onboard systems offer the fastest response times, while GCS/Cloud systems exhibit the highest latency due to communication overhead.
  • Energy: Onboard and swarm-cooperative approaches demand the highest computational resources from individual UAVs, whereas edge and cloud deployments offer significant offloading benefits.
  • Accuracy vs. Resources: There is a distinct trade-off where complex, high-accuracy models must be compressed (via pruning, quantization, or distillation) to fit onboard constraints, often necessitating a hybrid architecture where simple models run onboard and complex models run at the edge or cloud.
• Data Limitations: A significant finding is the reliance of many studies on generic IT datasets (e.g., KDD Cup 1999, NSL-KDD) which do not accurately reflect UAV-specific telemetry, bursty traffic, or RF impairments. This limits the ecological validity of reported results.

5. Significance and Claims

The paper claims its significance lies in providing a comprehensive, unified framework for UAV security that bridges the gap between cyber and physical domains.

• Holistic Defense: It argues that effective UAV security in smart cities cannot rely on siloed solutions; a unified IDS is necessary to correlate network anomalies with physical behaviors (e.g., linking a jamming event to a sudden deviation in flight path).
• Practical Guidance: By mapping detection techniques to specific deployment tiers (onboard vs. edge vs. cloud), the paper offers practical guidance for system architects facing resource constraints.
• Future Roadmap: The paper positions itself as a foundational guide for future research, emphasizing that the field must move beyond generic datasets toward UAV-specific benchmarks and address the "sim-to-real" gap. It highlights that while AI offers powerful tools, their deployment is currently hindered by data scarcity, computational limits, and the need for explainable, robust, and privacy-preserving mechanisms.

The authors conclude that while significant progress has been made, the realization of resilient, trustworthy smart cities depends on advancing these AI-driven IDS solutions to handle the complex, dual-nature threats of the UAV ecosystem.