Source-linked AI summary
Context-Aware Operational Security for Autonomous Drones
Burak Tufekci, Cihan Tunc
TL;DR
Drone security must contend with limited onboard resources and sequential operation data that conventional defenses may not adequately address. The paper proposes an LSTM-based IDS for drone sensor data and commands, reporting 98.08% classification accuracy with minimal latency in its evaluation.
Problem
Drone security must address sequential sensor data despite limited onboard resources, battery, mobility, and connectivity that can make conventional defenses insufficient.
Source: §1 · p. 2Method
The LSTM-based DUDE-IDS analyzes drone sensor data and received commands, with the authors implementing it on a physical drone for real-time detection.
Source: §1 · p. 2Results
98.08% classification accuracy while maintaining minimal latency was reported for DUDE-IDS in the drone evaluation.
Source: §5 · pp. 14, 18Takeaways & Limitations
The authors present DUDE-IDS as a practical real-time operational-security approach adaptable to MAVLink-using drones.
Source: §5 · p. 18Effectiveness on larger, more diverse datasets has not been fully explored.
Source: §7 · p. 20
Abstract
from arXiv · showhide
Autonomous drone-based services have been gaining significant interest across various application domains due to their mobility, flexibility, cost-effectiveness, and capability to integrate various sensors and actuators. However, operational failures or cyberattacks targeting drone systems can lead to severe economic impacts and safety concerns. Hence, ensuring secure and reliable autonomous drone operations is essential for the safe deployment of drone-enabled services. Nevertheless, the traditional security measures fall short due to drones' limited computational resources and power budget (battery), as well as the temporal and sequential behavior of drone operations due to drones' mobile nature. In this paper, we address this gap by utilizing Recurrent Neural Networks (RNNs), specifically focusing on Long Short-Term Memory (LSTM) networks, for autonomous drone operation security and reliability due to their temporal and sequential analysis capabilities. We leverage these capabilities for anomaly detection in autonomous drone sensor data and operation commands, which we refer to as Denial of Usage Detection Engine IDS (DUDE-IDS). We integrated the proposed DUDE-IDS into the drone mission computer (i.e., operating directly on drones rather than an edge node or ground control stations) to monitor data flows and to detect potential threats in real-time. Extensive experimental results demonstrate the effectiveness of this approach in identifying anomalies associated with GPS spoofing, Man-in-the-Middle (MITM), replay, and Denial-of-Service (DoS) attacks with 98% accuracy. We also evaluate our resource utilization and power consumption under different configurations, demonstrating the applicability of our approach in active drone operations in real-time.
1 Introduction
Autonomous drone missions need security that accounts for operational threats and the limits of conventional defenses on mobile, resource-constrained platforms. The paper motivates sequential-data analysis with RNNs and outlines a state-based, LSTM IDS, dataset, and physical-drone evaluation.
1 Introduction
- Drone operations face threats including sensor manipulation and control takeover, making security essential for safe and reliable missions.GNSS signal interception can disrupt operations, and reported incidents include compromised firmware and supply chains.Source: §1 · pp. 1–2
- Conventional defenses can be insufficient on drones with limited compute and battery, mobility, and connectivity.The paper highlights sequential sensor data as a basis for detecting anomalies and considering fail-safe responses to spoofing.Source: §1 · p. 2
- RNNs are proposed for anomaly detection because they retain context across sequences, unlike approaches that treat inputs independently or demand extensive resources.The introduction contrasts RNNs with feedforward networks, decision trees, CNNs, and transformers for temporal analysis and onboard use.Source: §1 · p. 2
- The paper defines drone operating states, creates a labeled MAVLink attack dataset, and proposes LSTM-based DUDE-IDS for sensor data and received commands.It also reports physical-drone real-time evaluation and comparison with a CNN across hardware environments; the dataset is intended for sharing.Source: §1 · p. 2
2 Related Work
Drone operations combine sequential sensor readings with time-ordered MAVLink commands, so deviations can signal hazards or failures. Prior work examines drone threats and defenses, while the paper identifies limited attention to MAVLink command-and-sensor monitoring on physical drones and limited public operational-control data.
2.1 Sequential Data in Drones
- Drone sensors and MAVLink commands form distinct sequential sources; deviations from expected command paths or sensor patterns may indicate attacks, faults, or hazards.MAVLink carries commands such as flight-mode changes, waypoints, parameter adjustments, and mission instructions.Source: §2.1 · p. 3
2.2 Cyber-Threats Targeting Drone Operations
- Prior work covers drone communication threat taxonomies and GPS-spoofing defenses, alongside MITM, replay, flooding, and cyber-physical IDS approaches.A cyber-physical IDS study reported improved detection from fusing cyber and physical data.Source: §2.2 · pp. 3–4
- The authors identify limited research on MAVLink command-and-sensor monitoring for operational safety, few physical-drone real-time evaluations, and scarce public control datasets.Their focus includes sensor spoofing, MITM, replay, and DoS attacks.Source: §2.2 · pp. 4–5
3 Methodology
DUDE-IDS runs on the mission computer to monitor MAVLink traffic and detect anomalies using drone operational context. It converts message features into sliding windows for many-to-one LSTM classification.
3 Methodology
- DUDE-IDS runs on the mission computer to monitor MC-FC command traffic, drop anomalous commands, and send safety actions to the flight controller.A queue stores and initially processes telemetry and controller data to prevent performance bottlenecks.Source: §3 · p. 5
3.1 Operational States of Drones
- The state machine models fundamental operations and routes abnormal commands or attacks to Error, with options to resume Mission, Land, or Idle.For GPS spoofing during Go to Waypoint, the drone can enter Error and return to Mission for Return to Base.Source: §3.1 · pp. 6–7
3.2 Feature Extraction for Drone Operation Analysis
- DUDE-IDS evaluates MAVLink command sequences against state transitions and combines command, sensor, and actuator readings to identify anomalies.Features include message direction, length, sequence number, source, component and message IDs, compatibility flags, current and previous states, and command parameters.Source: §3.2 · pp. 7–8
3.3 Feature Encoding and Sequence Construction
- Each MAVLink message becomes a fixed-dimensional vector combining encoded protocol/header features, current and previous state context, and normalized COMMAND_LONG parameters.Missing command parameters are zero-filled with a mask bit to preserve a fixed-dimensional representation.Source: §3.3 · p. 9
- The LSTM classifies each sliding window of L message-level timesteps as Normal, GPS Spoofing, MITM, Replay, or DoS.Stride balances detection latency and computational cost; windows shorter than L are zero-padded with mask bits.Source: §3.3 · p. 10
3.4 LSTM-based Anomaly Detection for DUDE-IDS
- LSTM gates selectively incorporate new observations and retain prior cell-state context, helping distinguish transient noise from persistent deviations.The input gate regulates new information, while the forget gate controls retention of the previous cell state.Source: §3.4 · pp. 10–11
- The final DUDE-IDS pipeline passes the LSTM's final hidden state through a fully connected layer and softmax to predict the window's class.Figure 4 depicts the many-to-one classifier operating on the sliding-window tensor.Source: §3.4 · p. 11, Figure 4 · p. 12
4 Experimental Setup
The experimental setup combines a custom drone testbed with labeled normal and attack flights, onboard monitoring, and LSTM/CNN models trained on extracted features. It evaluates scenarios repeatedly and shares the dataset materials for reproducibility.
4.1 Experimental Testbed
- The testbed paired a Raspberry Pi 4 Model B mission computer with a Pixhawk 2.4.8 flight controller, and used radio and SDR tools to stage attacks.The equipment included a 3DR Radio, HackRF One, Universal Radio Hacker, and GPS-SDR-SIM.Source: §4.1 · p. 12
4.2 Dataset Generation
- MAVLink logs were timestamped and labeled by scenario, with windows spanning multiple labels discarded to avoid ambiguous supervision.The dataset includes raw logs, synchronized sensor streams, scenario timestamps, and scripts for window extraction and feature encoding, to be shared publicly.Source: §4.2.1 · p. 12
4.2.1 Dataset Summary and Labeling.
- An onboard Python process continuously logged telemetry, sensor readings, actuator commands, and MAVLink commands, with minimal overhead intended to preserve drone performance.The dataset summary is identified in Table 5, but the supplied caption gives no cell values.Source: §4.2.2 · pp. 12–13
4.2.2 Monitoring Module Setup.
- Normal flights followed predefined paths, while cyberattack scenarios manipulated GPS, replayed prior data, flooded communications, or injected incorrect commands mid-flight.Each scenario was executed ten times to account for sensor noise and environmental factors.Source: §4.2.3 · p. 13
4.2.3 Flight Scenarios.
- Features were standardized and split into 80% training and 20% testing; the LSTM used three layers and 50 hidden units, while a CNN was implemented for comparison.The selected learning rate of 0.001 achieved 98.08% test accuracy and 0.9713 F1; 0.01 produced perfect test metrics but was described as overfitting, while 0.0001 learned poorly.Source: §4.3 · pp. 13–14
4.3 LSTM and CNN Model Generation
- Training used cross-entropy and Adam for ten epochs; four or more epochs reached 98.08% test accuracy and F1 stabilized around 0.9713.Two or three epochs yielded 80.77% test accuracy and 0.7218 F1; the CNN used two convolutional layers with 16 and 32 filters and a 1×3 kernel.Source: §4.3 · p. 14
5 Experimental Results
On the Raspberry Pi test platform, LSTM DUDE-IDS outperformed the CNN in classification, while the CNN had lower inference latency. The experiments also report platform-specific power and compute use, and assess runtime feasibility with latency and throughput measures.
5 Experimental Results
- 98.08% LSTM accuracy exceeded the CNN’s 96.45% on the test dataset.LSTM precision, recall, and F1 were 0.9620, 0.9808, and 0.9713; CNN values were 0.9302, 0.9645, and 0.9470.Source: §5 · pp. 14–15, Table 7 · p. 15
5.1 Baselines and Runtime Evaluation
- 250–300 ms overall detection delay supports real-time or near-real-time operation; LSTM inference was below 25 ms per window versus approximately 12 ms for CNN.With a window size of 20 messages, effective latency includes both window collection and inference; CNN latency was lower, while LSTM achieved higher accuracy.Source: §5 · p. 15
5 Experimental Results
- On Raspberry Pi, baseline power was around 2.75 W, compared with about 4.7 W for YOLO V8 and 5.3 W for DUDE-IDS idle; DoS averaged around 6.0 W and peaked at 6.5 W.DUDE-IDS used 3.5 of the Raspberry Pi’s four cores during operation.Source: §5 · p. 16, Table 9 · p. 16
- On Jetson Nano, baseline power was around 3.22 W and DUDE-IDS idle use was approximately 4.54 W; attack scenarios used 4.50–4.75 W, while DoS averaged 4.53 W and peaked at 5.25 W.DUDE-IDS used one of four Jetson Nano cores; the supplied text says this platform handled attacks with less variability.Source: §5 · pp. 16–17, Table 10 · p. 17
5.1.1 Baselines.
- The runtime evaluation includes rule-based, Random Forest, and Normal-only One-Class SVM/Isolation Forest baselines; Jetson Nano provided higher throughput and lower inference time per window.The study measured latency per window and sustained throughput; the supplied Table 11 caption identifies runtime statistics but provides no cell values.Source: §5.1.2 · p. 19, §5.1.1 · p. 19, Table 11 · p. 19
6 Further Discussions
The discussion describes DUDE-IDS’s portability boundaries and how environmental variation and changing attack patterns shape deployment and future hardening.
Scalability to diverse drone platforms
- MAVLink semantics and platform-agnostic states support portability, but the study’s custom drone and finite scenarios leave generalization across airframes and missions unquantified.MAVLink platforms need minimal modification; non-MAVLink systems require equivalent command vocabularies and state abstractions.Source: §6 · p. 19
Environmental variability and sensor noise
- Wind, GPS multipath, RF interference, and payload dynamics can shift feature distributions and increase false alarms; multi-sensor context and sequence modeling are proposed mitigations.Mission-specific threshold calibration and periodic recalibration flights can update normalization statistics.Source: §6 · pp. 19–20
Evolving attack patterns and adversarial adaptation
- Replay cadence changes, benign-command blending, and low-and-slow drift remain detectable when they create inconsistent cross-feature dynamics over time.Proposed hardening includes drift detection, fine-tuning on recent benign logs, and lightweight ensembles to reduce false positives without sacrificing sensitivity.Source: §6 · p. 20
7 Conclusion
The conclusion reports strong LSTM anomaly-detection performance and evaluates deployment resources, while noting limits in dataset scale, device demands, and cost.
Conclusion
- 98.08% classification accuracy for LSTM exceeded the CNN model’s 96.45%, with LSTM precision 0.9620, recall 0.9808, and F1 score 0.9713.The authors report these results for DUDE-IDS on drone sensor data and commands.Source: §7 · p. 20
- Jetson Nano peaked at around 5.25 W and used 1 of 4 cores, while Raspberry Pi peaked at 6.5 W during DoS attacks and used 3.5 of 4 cores.Both platforms handled DUDE-IDS with minimal latency; the authors attribute Jetson’s efficiency to its dedicated GPU and hardware optimization for ML/AI tasks.Source: §7 · p. 20
- Larger, more diverse datasets have not been fully explored, Raspberry Pi demands may limit long-term feasibility for complex tasks, and Jetson Nano’s higher cost may hinder large-scale deployment.These are stated limitations of the study and its deployment options.Source: §7 · p. 20
- The authors conclude that DUDE-IDS is a promising ML-based framework for improving drone cybersecurity and operational reliability against varied attack vectors.The conclusion emphasizes LSTM analysis of sequential drone sensor data and commands.Source: §7 · p. 20