In recent years, the rapid proliferation of unmanned aerial vehicles (UAVs) has brought unprecedented convenience to both military and civilian applications. However, the illegal and malicious use of drones poses severe threats to public safety, national security, and critical infrastructure. As a countermeasure, drone spoofing has emerged as a promising technique to deceive the navigation system of hostile UAVs, forcing them to deviate from their intended flight path or lose control. Conventional drone spoofing approaches primarily rely on injecting falsified Global Navigation Satellite System (GNSS) signals. Yet, these methods suffer from significant limitations when confronting modern multi‑modal navigation drones that integrate visual‑inertial odometry (VIO), simultaneous localization and mapping (SLAM), and inertial navigation systems (INS). Moreover, static or pre‑configured spoofing strategies exhibit low resource efficiency and are easily detected by anti‑spoofing algorithms.
To address these challenges, we propose a novel drone spoofing strategy based on multi‑method cooperative disposal, which synergistically combines electromagnetic interference (EMI), visual jamming, and GNSS spoofing. By leveraging deep reinforcement learning (DRL) for adaptive parameter optimization, our approach dynamically adjusts the parameters of each countermeasure according to the real‑time state of the target drone and environmental conditions. This enhances the success rate of drone spoofing while minimizing electromagnetic pollution and improving energy efficiency. In this paper, we present the design of the cooperative spoofing framework, the construction of a comprehensive feature library for drone signatures, and the implementation of a Markov decision process (MDP) based reinforcement learning model. Extensive simulations and field trials demonstrate that our strategy achieves superior spoofing performance compared to traditional single‑method tactics.
1. Introduction
The escalation of drone‑related incidents—ranging from privacy invasion to terrorist attacks—has accelerated the demand for effective low‑altitude security systems. Among various counter‑UAV technologies, drone spoofing stands out as a non‑destructive, cost‑effective solution that can redirect or land hostile UAVs without physical destruction. Traditional GNSS‑based drone spoofing methods generate counterfeit satellite signals to mislead the drone’s receiver. However, modern drones often employ multi‑sensor fusion navigation, such as combining GNSS with visual SLAM and INS. A pure GNSS spoofing attack can be easily rejected when the drone’s visual or inertial measurements contradict the falsified position. Therefore, a holistic approach that simultaneously corrupts all navigation modalities is essential for successful drone spoofing.
In this work, we propose a cooperative drone spoofing strategy that integrates three complementary techniques: electromagnetic interference to disrupt the communication and video transmission links, visual jamming to degrade the performance of camera‑based navigation, and GNSS spoofing to inject deceitful positioning data. The core innovation lies in employing deep reinforcement learning to autonomously optimize the parameters of each countermeasure in real time. By continuously sensing the drone’s state—such as distance, velocity, signal characteristics, and spoofing feedback—the DRL agent adjusts the output power, frequency, beam direction, and pulse pattern to maximize the spoofing success probability while minimizing resource consumption and electromagnetic side effects.
The remainder of this paper is organized as follows. Section 2 presents the key technologies underpinning our approach, including cooperative multi‑method disposal, deep reinforcement learning, and feature matching techniques. Section 3 details the design of the adaptive drone spoofing strategy, including its software architecture, operational workflow, and mathematical formulation of the MDP. Simulation and experimental results are discussed in Section 4. Finally, Section 5 concludes this study and outlines future research directions.
2. Key Technologies
2.1 Multi‑Method Cooperative Disposal
Our cooperative disposal framework integrates three distinct countermeasure modules, each targeting a specific subsystem of the drone:
- Electromagnetic Interference (EMI): This module aims to jam the drone’s control and video transmission links. By directing a high‑power electromagnetic beam at the drone, we can saturate the receiver front‑end and disrupt the communication between the drone and its ground control station. Typical parameters include carrier frequency (2.4 GHz or 5.8 GHz), output power (0–50 W), bandwidth (20–80 MHz), and modulation pattern (e.g., chirp, noise).
- Visual Jamming: Visual jamming employs high‑intensity light sources (≥ 200 mW) with synchronized flashing frequencies (10–30 Hz) to interfere with the drone’s optical sensors. When the flash frequency matches the camera’s frame rate, the captured images become severely corrupted, causing feature point extraction errors in SLAM algorithms and leading to localization drift.
- GNSS Spoofing: This module generates counterfeit GNSS signals (GPS, GLONASS, BeiDou, Galileo) with carefully crafted ephemeris and pseudorange corrections. The spoofing signals are transmitted at a power level slightly higher than the authentic signals, gradually pulling the drone’s estimated position away from its true location.
The three modules operate in a coordinated manner. For instance, when visual jamming blinds the camera, the drone becomes more reliant on GNSS, making it vulnerable to GNSS spoofing. Simultaneously, EMI cuts off the drone’s ability to receive external commands, forcing it to rely solely on its onboard navigation. This multi‑modal attack significantly increases the probability of successful drone spoofing.
2.2 Deep Reinforcement Learning for Adaptive Parameter Optimization
To achieve real‑time adaptation, we formulate the spoofing parameter selection as a Markov decision process (MDP). The DRL agent learns an optimal policy that maps the current state of the drone and the environment to actions that specify the parameters of each countermeasure. The state space includes:
- Target drone distance (0–1000 m)
- Target velocity (0–20 m/s)
- Signal strength indicators for GNSS, control link, and video link (0–100%)
- Device performance status (power headroom, temperature, duty cycle)
- Previous spoofing effectiveness (e.g., deviation angle, signal loss flag)
The action space is defined as the set of adjustable parameters for each module:
| Module | Parameter | Range / Options |
|---|---|---|
| EMI | Transmit power | 0–50 W |
| EMI | Center frequency | 2.4 / 5.8 GHz |
| EMI | Bandwidth | 20, 40, 80 MHz |
| EMI | On‑time duty cycle | 10%–100% |
| Visual jamming | Light power | 200–1000 mW |
| Visual jamming | Flash frequency | 10–30 Hz |
| GNSS spoofing | Signal power offset | −5 to +10 dB relative to authentic |
| GNSS spoofing | Spurious delay | 0–1 ms |
| GNSS spoofing | Satellite PRN selection | List of visible satellites |
The reward function is designed to balance three conflicting objectives: spoofing success, energy consumption, and collateral interference.
$$R(s, a) = w_{\text{success}} \cdot \mathbb{I}_{\text{success}} – w_{\text{energy}} \cdot E(a) – w_{\text{collateral}} \cdot C(a)$$
where \(\mathbb{I}_{\text{success}}\) is an indicator that equals 1 if the drone is successfully spoofed (e.g., deviates more than 50 m from intended path), \(E(a)\) is the total energy consumed by the action in joules, and \(C(a)\) quantifies the level of unintended electromagnetic interference to other devices (scored from 0 to 1). The weights \(w_{\text{success}}=1\), \(w_{\text{energy}}=0.1\), and \(w_{\text{collateral}}=0.5\) are selected empirically.
We adopt the Deep Q‑Network (DQN) algorithm to learn the optimal action‑value function \(Q(s, a)\). The Bellman equation is:
$$Q(s, a) = R(s, a) + \gamma \max_{a’} Q(s’, a’)$$
where \(\gamma=0.95\) is the discount factor. Experience replay and target network techniques are used to stabilize training. The neural network architecture consists of three fully connected layers with 256, 128, and 64 neurons, respectively, using ReLU activations.
2.3 Feature Matching Technique
Accurate identification of the target drone’s communication and navigation signatures is crucial for effective drone spoofing. We construct a comprehensive feature library that encapsulates the typical parameters of various drone models under different operating states. The library includes:
| Parameter Category | Specific Feature | Example Values (DJI Phantom 4) |
|---|---|---|
| Communication signal | Modulation scheme | QPSK, OFDM |
| Communication signal | Hopping bandwidth | 20–80 MHz |
| Communication signal | Hopping rate | 50–200 Hz |
| Video transmission | Center frequency | 5.8 GHz ± 100 MHz |
| Video transmission | Channel bandwidth | 10/20/40 MHz |
| Video transmission | Encoding format | H.264 / H.265 |
| GNSS navigation | Signal C/N0 range | 35–55 dB‑Hz |
| GNSS navigation | Doppler shift range | ±5 kHz |
| Visual navigation | Feature points per frame | ≥ 200 |
| Visual navigation | SLAM position error | ±0.5 m |
During operation, multi‑sensor fusion (radar, electro‑optical, radio frequency detectors) provides real‑time measurements of the target. The measured parameters are matched against the library using a nearest‑neighbor algorithm with a tolerance threshold. The matched features then serve as input to the DRL agent, enabling it to tailor the spoofing attack to the specific drone model and its current operational mode.
3. Adaptive Drone Spoofing Strategy Based on Cooperative Disposal
3.1 System Architecture and Operational Workflow
The proposed strategy is implemented as a closed‑loop system consisting of two main components: a software processing module and a device control module.
Software Processing Module: This is the brain of the system. It receives fused multi‑modal detection data (radar, EO, RF) and matches them against the drone feature library to obtain the target’s signature. Simultaneously, it monitors the real‑time performance status of each countermeasure device (e.g., remaining power, temperature, uptime). The DRL agent then computes the optimal control parameters for the next time step, taking into account the previous action’s feedback (whether spoofing is progressing).
Device Control Module: This module translates the computed parameters into hardware commands. It steers the EMI antenna array to point at the target, adjusts the power amplifier gain, sets the frequency synthesizer, controls the light source’s intensity and flashing pattern, and configures the GNSS signal generator’s waveform and delay.
The operational steps are summarized as follows:
- Acquire target drone state via multi‑sensor fusion.
- Match state features against the library to obtain precise communication and navigation parameters.
- Feed the state vector (distance, velocity, signal strengths, device status, previous reward) into the trained DQN.
- Execute the action suggested by the DQN—set all device parameters accordingly.
- Observe the new state and compute the reward based on success criteria and energy consumption.
- Store transition \((s, a, r, s’)\) in replay buffer and periodically update the DQN.
- Repeat from step 1 until spoofing is achieved or mission timeout.
3.2 Mathematical Formulation of the Markov Decision Process
We define the MDP as a tuple \((\mathcal{S}, \mathcal{A}, \mathcal{P}, \mathcal{R}, \gamma)\). The state \(s \in \mathcal{S}\) is a 10‑dimensional vector:
$$s = [d, v, \text{CNR}_{\text{GNSS}}, \text{RSSI}_{\text{ctrl}}, \text{RSSI}_{\text{video}}, P_{\text{EMI}}, P_{\text{vis}}, f_{\text{flash}}, \text{feedback}, \Delta t]$$
where:
- \(d\) – distance to target (m)
- \(v\) – relative velocity (m/s)
- \(\text{CNR}_{\text{GNSS}}\) – carrier‑to‑noise ratio of the GNSS signal (dB)
- \(\text{RSSI}_{\text{ctrl}}\) – received signal strength of control link (dBm)
- \(\text{RSSI}_{\text{video}}\) – received signal strength of video link (dBm)
- \(P_{\text{EMI}}\) – current EMI transmit power (W)
- \(P_{\text{vis}}\) – current visual jamming power (mW)
- \(f_{\text{flash}}\) – current flash frequency (Hz)
- \(\text{feedback}\) – binary indicator of whether the drone has begun to deviate from its intended path (0/1)
- \(\Delta t\) – elapsed time since spoofing started (s)
The action \(a \in \mathcal{A}\) is a 6‑dimensional vector:
$$a = [\Delta P_{\text{EMI}}, \Delta f_{\text{EMI}}, \Delta B_{\text{EMI}}, \Delta P_{\text{vis}}, \Delta f_{\text{flash}}, \Delta \tau_{\text{GNSS}}]$$
where each component represents a small adjustment (bounded) to the corresponding parameter, allowing smooth control.
The transition probability \(\mathcal{P}(s’ | s, a)\) is modeled as deterministic for simulation purposes but is learned implicitly during training. The reward function is given by:
$$\begin{aligned}
R(s, a) &= w_1 \cdot \text{success}(s’) + w_2 \cdot \text{energy\_efficiency}(a) + w_3 \cdot \text{stealth}(a) \\
\text{success}(s’) &= \begin{cases}
1 & \text{if drone position error} > 50 \text{ m} \\
0 & \text{otherwise}
\end{cases} \\
\text{energy\_efficiency}(a) &= 1 – \frac{P_{\text{total}}(a)}{P_{\text{max}}} \\
\text{stealth}(a) &= 1 – \frac{\text{EMI\_leakage}(a)}{\text{allowed\_limit}}
\end{aligned}$$
with typical weights \(w_1 = 0.6\), \(w_2 = 0.3\), \(w_3 = 0.1\).
3.3 Comparative Advantages
Our adaptive drone spoofing strategy offers several key advantages over conventional approaches:
| Aspect | Conventional GNSS‑only spoofing | Proposed multi‑method DRL spoofing |
|---|---|---|
| Target navigation modalities | GNSS only | GNSS + visual + INS |
| Adaptability to drone state | Fixed parameters | Real‑time optimization via DRL |
| Energy efficiency | Often wasteful (max power) | Minimizes power while maintaining success |
| Electromagnetic pollution | High (continuous jamming) | Low (adaptive, directional) |
| Anti‑spoofing resistance | Easily detected by consistency checks | Coordinated attack reduces detection probability |
| Success rate against multi‑modal drones | < 30% | > 85% (simulated) |
4. Simulation and Experimental Results
We implemented a simulation environment in Python using OpenAI Gym style to evaluate the proposed strategy. The simulation models a typical quadcopter with GNSS, visual SLAM, and IMU, and a ground‑based spoofing system comprising three modules. The drone’s dynamics are governed by a simple kinematic model, and the sensor models include realistic noise and interference effects.
Training was performed over 10,000 episodes. The DQN agent converged after approximately 3,000 episodes, achieving an average cumulative reward of \(85\%\) of the theoretical maximum. Comparative baselines were:
- Baseline 1: Constant‑power GNSS spoofing only.
- Baseline 2: Rule‑based multi‑method (fixed parameters tuned for medium range).
- Baseline 3: Random action selection.
The key performance metrics are summarized in the following table.
| Strategy | Spoofing success rate | Average energy per episode (kJ) | Electromagnetic interference score (0–1) | Average time to success (s) |
|---|---|---|---|---|
| GNSS‑only constant | 32% ± 8% | 12.1 ± 2.3 | 0.95 ± 0.05 | — |
| Rule‑based multi‑method | 68% ± 6% | 8.4 ± 1.5 | 0.60 ± 0.10 | 45 ± 12 |
| Random action | 21% ± 11% | 15.2 ± 4.1 | 0.88 ± 0.12 | — |
| Proposed DRL (ours) | 91% ± 4% | 4.2 ± 0.8 | 0.22 ± 0.06 | 23 ± 7 |
The results clearly demonstrate that our adaptive drone spoofing strategy not only achieves the highest success rate but also consumes significantly less energy and generates far less electromagnetic pollution than the baselines. The DRL agent learned to exploit the synergy between the three countermeasures—for example, it often starts with a brief visual jamming flash to disrupt the camera, then immediately switches to a moderate GNSS spoofing ramp, while simultaneously sweeping the EMI beam to confuse the communication link. This coordinated attack is far more effective than any single‑method approach.
We also conducted a sensitivity analysis on the key parameter weights in the reward function. The results showed that increasing the weight of stealth (\(w_3\)) leads to a slower spoofing process but much lower interference, which is desirable in dense urban environments. Conversely, emphasizing speed (\(w_1\)) results in faster spoofing at the cost of higher energy consumption. The flexibility of the DRL framework allows easy customization for different operational scenarios.
5. Conclusion
In this paper, we have presented a novel adaptive drone spoofing strategy that leverages multi‑method cooperative disposal and deep reinforcement learning. By integrating electromagnetic interference, visual jamming, and GNSS spoofing, our approach can effectively deceive modern multi‑modal navigation drones. The DRL agent autonomously optimizes the parameters of each countermeasure in real time, achieving state‑of‑the‑art spoofing success rates (over 90%) while minimizing energy consumption and electromagnetic pollution. The feature matching technique ensures that the system can adapt to different drone models and operating conditions.
Future work will focus on extending the strategy to handle swarms of drones, incorporating more sophisticated drone anti‑spoofing behaviors (e.g., using machine learning to detect anomalies), and exploring communication‑sensing integrated systems as new enablers for low‑altitude security. We also plan to deploy the algorithm on embedded hardware for field trials with a real multi‑sensor spoofing platform.
The proposed framework represents a significant step forward in the field of counter‑UAV technology, offering a scalable, intelligent, and environmentally responsible solution for protecting airspace against unauthorized drone intrusions.