Research on Coordinating UAV Communication Networks to Suppress Co-channel Interference Based on Spatio-Temporal Graph Neural Networks
DOI:https://doi-001.org/1025/17745173089796
Yiyang Chen1*
1University of Science and Technology Beijing, Beijing, China
E-mail: 564651692@163.com
Abstract: The high-density deployment of drone swarms in emergency communication and wide-area monitoring applications faces significant performance constraints from co-channel interference. Existing approaches relying on static topology optimization or centralized architectures struggle to address dynamic interference coupling in mobile scenarios. This study proposes a collaborative anti-interference framework integrating spatiotemporal graph neural networks with distributed intelligence. The methodology incorporates dynamic interference graph modeling, federated reinforcement learning for distributed coordination, and joint hardware-algorithm optimization incorporating intelligent reflecting surfaces. Experimental results demonstrate substantial improvements in network throughput and spectral efficiency while maintaining millisecond-level response latency in dense deployment scenarios. The framework achieves effective balance between interference suppression accuracy and computational overhead through its hierarchical architecture combining global parameter aggregation with local decision-making. This research provides theoretical foundations for reliable drone networking in dynamic electromagnetic environments and contributes to the paradigm shift from static optimization to adaptive interference management. The proposed approach offers technical pathways for 6G air-ground integrated networks, with future work focusing on cross-layer optimization and edge deployment for smart city and emergency response applications.
Keywords: Spatiotemporal graph neural network, Unmanned aerial vehicle (UAV), Co-channel interference suppression, Federated learning, Intelligent reflecting surface, Integration of communication and computing
1. Introduction
With the rapid development and large-scale application of drone technology, drone swarms have demonstrated immense potential in fields such as emergency communication, wide-area monitoring, and intelligent inspection (Khargharia et al., 2025). As the core infrastructure supporting their collaborative operations, drone communication networks must achieve high reliability and low latency data transmission in complex electromagnetic environments (Li et al., 2025; Qiu & Ji, 2025). However, with the increasing deployment density of drones and the growing complexity of tasks, the issue of co-channel interference has become increasingly prominent, posing a key bottleneck that constrains network performance. In dense networking scenarios, competition for spectrum resources among drone nodes intensifies, leading to frequent issues such as communication quality degradation and link interruptions caused by signal overlap (Singh et al., 2025; Banjar & Alshdadi, 2025). Traditional interference management methods based on fixed topology struggle to adapt to the dynamic characteristics of drone networks. The topological variability and channel state uncertainty triggered by high-speed node movement further increase the difficulty of interference suppression (Zerrougui & Calagate, 2025). How to achieve accurate interference modeling and real-time coordinated control in dynamic environments has become a pressing technical challenge in the field of drone communication.
Existing research on addressing co-channel interference in drone networks primarily focuses on single-dimensional optimizations such as physical layer beam forming or network layer resource allocation, lacking a systematic exploration of spatiotemporal correlation characteristics (Liao et al., 2025). Although traditional graph neural networks can model the spatial dependencies between nodes, they struggle to capture the temporal evolution laws of multi-dimensional parameters such as drone trajectories and channel states (Zhou et al., 2025). Furthermore, existing cooperative anti-interference mechanisms generally adopt centralized decision-making architectures, which are difficult to meet the stringent requirements of real-time performance and scalability in large-scale networks. As the number of network nodes increases, the random distribution of interference sources, dynamic competition for spectrum resources, and exponential growth in computational load impose complex constraints on interference suppression technology requiring multi-dimensional coupling (Hou & Wang, 2026; Bang & Kim, 2025). How to construct a new anti-interference framework that balances spatiotemporal feature perception and distributed cooperative optimization has become a core challenge in enhancing the robustness and spectrum efficiency of drone networks.
This paper proposes a collaborative anti-interference method for unmanned aerial vehicle (UAV) communication networks based on spatiotemporal graph neural networks to address the aforementioned issues. By deeply integrating the spatiotemporal motion characteristics of UAV nodes with channel state information, a dynamically evolving network interference graph is constructed to achieve joint optimization of interference source localization and spectrum conflict prediction. A hierarchical collaborative architecture is designed, combining global federated learning with local real-time decision-making, to enhance algorithm execution efficiency while ensuring interference suppression accuracy. This method breaks through the limitations of traditional static models, provides new ideas for spectrum resource management in dynamic topology environments, and has important theoretical value for promoting reliable networking of UAV swarms in complex scenarios. The subsequent chapters will delve into system modeling, algorithm design, experimental verification, and other dimensions, systematically elaborating on the implementation path and innovative breakthroughs of the technical solution.
2. Theoretical framework of collaborative anti-interference based on spatiotemporal graph neural network
2.1 Application of spatiotemporal graph neural networks in communication networks
In recent years, graph neural networks have been widely applied in the field of communication network optimization due to their powerful relationship modeling capabilities (Chen & Li, 2025). Early research primarily relied on static graph structures, mapping network nodes and links to a fixed topology, and extracting spatial features through graph convolution to support resource allocation decisions (Su et al., 2025; Su & Cai et al., 2025; Komarovsky & haddad, 2025). However, the dynamic nature of drone networks leads to continuous changes in node positions, channel states, and interference relationships, making it difficult for traditional static graph models to capture real-time evolutionary patterns (Srivastava et al., 2025). To address this, scholars have proposed spatiotemporal graph neural networks combined with time series analysis, incorporating recurrent neural networks or attention mechanisms to jointly model topological temporal features. For instance, some works utilize long short-term memory networks to predict node movement trajectories and update the dynamic adjacency matrix through graph convolution, significantly enhancing the accuracy of spectrum allocation (Zhang et al., 2025; Qin et al., 2024). Nevertheless, existing methods still face challenges such as high computational complexity and insufficient real-time performance, especially in dense drone scenarios where frequent topological changes significantly reduce model training and inference efficiency.
2.2 Co-channel interference suppression technology in drone communication networks
The suppression of co-channel interference in drone communication is a core issue in ensuring network performance. Early research focused on physical layer technologies, such as beam forming and power control, to reduce the impact of interference by optimizing the direction and intensity of signal transmission (Ayegun et al., 2025; Lv et al., 2025). These methods rely on precise channel state information, but in high-speed drone mobility scenarios, channel measurement errors and feedback delays lead to a sharp deterioration in performance (Zhang et al., 2025). In recent years, interference suppression technologies based on intelligent reflecting surfaces have gained attention. They attenuate the strength of interference signals by dynamically adjusting the propagation paths of electromagnetic waves. However, these technologies come with high hardware costs and limited deployment flexibility (Wang et al., 2025). Network layer methods, on the other hand, focus on spectrum resource scheduling, employing game theory or reinforcement learning strategies to achieve distributed decision-making. However, existing algorithms generally ignore the coupling relationship between spatial and temporal dimensions, making it difficult to balance interference suppression and communication efficiency in dynamic environments (Liao et al., 2024). For example, some studies allocate spectrum resources through fixed rules, which can reduce conflicts but lead to low spectrum utilization.
2.3 Research progress on collaborative anti-jamming technology
Collaborative anti-jamming technology, which enhances the overall network performance through joint optimization among multiple nodes, has become a research hot spot in recent years. Centralized methods rely on global information for unified decision-making, achieving theoretical optimality, but their practical application is limited by communication overhead and computational delay (Gao et al., 2025). Distributed methods allow nodes to make autonomous decisions based on local information, but the lack of a global perspective can lead to local optima (Jin & Shi, 2023). Compromise solutions, such as hierarchical collaborative architectures, aggregate local model updates through federated learning to update global strategies, reducing communication costs while improving decision consistency (Zhagn & Wen, 2023; Du & Jia, 2023). Some works further introduce graph neural networks to model and optimize collaboration efficiency using node relationships, but the model’s generalization ability in dynamic environments remains to be verified (Zhao & Liu et al., 2022). It is worth noting that existing collaborative mechanisms are mostly designed for static or low-dynamic scenarios, lacking adaptability to the rapid topological changes and complex interference coupling in drone networks. There is an urgent need to explore new collaborative paradigms that balance real-time performance and accuracy.
As shown in Figure 1, based on spatiotemporal graph neural networks, this paper proposes a hierarchical collaborative anti-interference method for unmanned aerial vehicles (UAVs). Firstly, a dynamic interference graph is constructed to accurately model the spatiotemporal correlation of spectrum conflicts. Then, a hierarchical decision-making mechanism integrating federated learning and reinforcement learning is designed to achieve collaborative optimization of interference suppression and resource allocation. Furthermore, intelligent reflecting surfaces are introduced to assist beam forming, and the physical layer’s anti-interference capability is enhanced through joint design of hardware algorithms. Experiments show that the proposed method significantly outperforms traditional schemes in terms of throughput, latency, and interference suppression, providing an effective solution for reliable communication under dynamic topology.
Figure 1. Article structure diagram
3. Dynamic interference spectrum modeling and collaborative optimization method
3.1 System model and problem modeling
The dynamic nature and interference complexity of drone communication networks necessitate the establishment of an accurate mathematical description framework (Banjar & Alshdadi, 2025). In dense networking scenarios, the spatiotemporal coupling characteristics of node locations, channel states, and interference relationships pose multidimensional challenges for modeling.
Figure 2. System architecture diagram
Existing research often adopts static topology assumptions, ignoring the dynamic evolution of the network caused by the high-speed movement of drones, leading to insufficient accuracy in interference prediction. To address this issue, this section proposes a method based on joint modeling of dynamic location and channel state. The three-dimensional motion trajectory of nodes is defined by formula (1), and the channel gain model is established by combining formula (2), accurately characterizing the path loss and phase variation characteristics of signal propagation. Furthermore, formula (3) is introduced to quantify the co-channel interference power, clarifying the direct impact of frequency band conflicts on communication quality. Finally, formula (4) constructs an optimization objective centered on the total network rate, incorporating power constraints and frequency band capacity limitations into a unified framework, laying the foundation for subsequent algorithm design (Figure 2). Drone location dynamic model:
(1)
: the three-dimensional coordinates of the drone at any given moment. Channel gain model:
(2)
Where, is the reference path loss; is the node spacing; is the path loss index; and is the phase offset. Co-channel interference power calculation.
(3)
In the formula, represents the transmit power, and represents the indicator function. The optimization objective is to maximize the total network rate.
(4)
: it refers to the noise power spectral density. The dynamic system model established in this section breaks through the limitations of traditional static assumptions. By integrating node kinematic parameters with channel state information, it achieves the first spatiotemporal joint modeling of interference coupling relationships, providing rigorous mathematical support for interference suppression in dynamic environments.
3.2 Spatio-temporal Graph Neural Network Architecture
The design of spatiotemporal graph neural networks requires consideration of both the dynamic nature of network topology and the heterogeneity of node features, posing dual challenges for feature extraction and relational reasoning (Chen & Li, 2025).
Figure 3. Spatio-temporal Graph Neural Network Structure Diagram
Traditional graph convolutional methods only focus on spatial dependencies, making it difficult to capture the temporal evolution of drone trajectories and channel states. To this end, as shown in Figure 3, this section proposes a hierarchical spatiotemporal feature fusion architecture. Formula (5) defines a multi-dimensional node feature matrix, encompassing location, velocity, power, and frequency band information; formula (6) dynamically calculates edge weights through a Gaussian kernel function and signal-to-interference-plus-noise ratio, accurately characterizing the interference correlation strength between nodes. On this basis, formula (7) employs Chebyshev polynomials to approximate graph convolution, significantly reducing computational complexity; formulas (8)-(11) design gated recurrent units (GRUs), achieving gradual fusion of temporal features through update and reset gate mechanisms. Node feature matrix:
(5)
Where is the velocity vector and is the frequency band number. Dynamic edge weight calculation:
(6)
: for Gaussian kernel width. Chebyshev graph convolution:
(7)
In the formula, represents the normalized Laplacian matrix; represents the Chebyshev polynomial of order ; represents the trainable parameter matrix. GRU update gate:
(8)
: to update the gate weight matrix, the GRU reset gate:
(9)
: to reset the gate weight matrix. Candidate state generation:
(10)
: it is the candidate state weight matrix. The final state update is as follows:
(11)
: it is a Hadamard product. This architecture achieves multi-granularity perception of network dynamic characteristics while ensuring computational efficiency through the collaborative design of spatial graph convolution and temporal recurrent units, providing high-dimensional feature representations for interference prediction and resource allocation.
3.3 Coordination mechanism and interference suppression strategy
Dynamic interference suppression requires balancing communication performance and resource consumption in multi-objective optimization, which poses stringent requirements on the design of collaborative mechanisms (Wang et al., 2024).
Figure 4. Schematic diagram of IRS-assisted beam forming and interference suppression
Existing power allocation and spectrum selection strategies often optimize independently, lacking joint consideration of spatiotemporal correlation. This section proposes a collaborative optimization framework based on reinforcement learning and attention mechanism. Formula (12) designs an adaptive power control strategy to dynamically adjust the transmit power according to the target signal-to-interference-plus-noise ratio; formula (13) introduces the Q-learning algorithm to iteratively optimize the spectrum selection strategy through the state-action value function. To further enhance the processing priority of high-interference links, formula (14) calculates link weights using the attention mechanism, and formulas (15)-(16) define the interference conflict intensity and spectrum switching cost models, respectively, forming the decision-making basis under multi-dimensional constraints (Figure 4). Dynamic power allocation:
(12)
: it is the power control coefficient. Q-learning action value update:
(13)
: it is the learning rate, which is the discount factor. Attention weight calculation:
(14)
: for query vectors and key vectors, the dimension of the vectors is specified. Interference conflict detection function:
(15)
: it represents the conflict intensity of nodes at time t. Spectrum switching cost model:
(16)
: to switch the penalty coefficient. This strategy achieves collaborative optimization of interference suppression and resource efficiency through the deep integration of reinforcement learning and attention mechanisms, providing a scalable solution for real-time decision-making in complex dynamic scenarios.
3.4 Optimization algorithm
The contradiction between global optimization and distributed execution is the core difficulty in collaborative optimization of drone networks, necessitating the design of an efficient and robust algorithm framework.
Traditional centralized optimization methods struggle to cope with the surge in computational load brought about by the expansion of network scale. As shown in Figure 5, this section proposes a distributed training framework empowered by federated learning. Formula (17) constructs a joint loss function that integrates communication rate, interference suppression, and model regularization objectives; formula (18) designs a federated aggregation rule that achieves distributed model parameter fusion through weighted averaging; and formula (19) employs a gradient descent algorithm to update model parameters, ensuring the convergence of the optimization process. Through alternating iterations of global federated learning and local model updates, this algorithm significantly enhances the generalization ability of interference suppression strategies while reducing communication overhead. Joint loss function:
(17)
For rate loss, for interference loss. Federal Aggregation Rules:
(18)
Where, is the size of the local dataset, and is the number of nodes participating in the federation. Gradient descent update:
(19)
: it refers to the learning rate. By combining hierarchical optimization architecture with federated learning mechanism, this algorithm effectively solves the collaborative optimization problem of large-scale dynamic networks, providing theoretical support for engineering deployment.
Figure 5. Flowchart of collaborative training in federated learning
The specific algorithm steps of this study are shown in Table 1.
Table 1. Algorithm Steps
| Step | Instructions | Core code example |
| 1 | Constructing a spatiotemporal graph structure | graph = build_spatiotemporal_graph(nodes) |
| 2 | Local model training | local_model.train(local_dataset, epochs=5) |
| 3 | Federated parameter aggregation | global_weights = fed_avg(local_weights_list) |
| 4 | Dynamic power allocation | power = min(max_power, eta * target_sinr / channel_gain) |
| 5 | Spectrum selection decision | action = q_network.argmax(state) |
| 6 | Calculate link priority | alpha = softmax(query @ key.T / sqrt(d)) |
| 7 | Real-time interference detection | if interference > threshold: switch_channel() |
| 8 | Gradient backpropagation | optimizer.step(loss.backward()) |
| 9 | Topology dynamic update | graph.update_edges(velocity, position) |
| 10 | Model deployment verification | deploy_model(global_weights, edge_device) |
This chapter establishes a dynamic modeling method for interference graph based on spatiotemporal feature fusion. The topological temporal evolution tracking is achieved through gated recurrent units (Equations 8-11). Experiments show that this method improves the prediction accuracy of spectrum conflicts by 38.6%. A hierarchical reinforcement learning architecture is proposed, combining Q-learning action value update (Equation 13) with attention priority scheduling (Equation 14), which reduces end-to-end delay by 53.3% in a 20-node scenario. This study is the first to jointly optimize the federated learning loss function (Equation 17) and IRS beam forming, verifying that this method can achieve a spectrum utilization rate of 95.4%, an improvement of 9.1% over the baseline method.
4. Experimental verification of interference suppression performance in multiple scenarios
4.1 Experimental setup
To verify the effectiveness of the proposed method, a multi-scenario dynamic drone communication network simulation environment was constructed for experiments. The simulation covers three typical scenarios: urban dense areas, mountainous areas, and maritime areas. The drone node scale dynamically expands from 4 to 20, and the ground terminals adopt a fixed deployment strategy. The communication frequency band is divided into 16 sub-bands ranging from 2.4 GHz to 5.8 GHz, and the transmission power is adaptively adjusted according to the channel state. The drone motion model adopts a Gaussian-Markov random process to simulate real flight trajectories, with the maximum speed set to a random distribution between 10-20 m/s. In terms of algorithm parameters, the spatiotemporal graph neural network updates its topological features every 50 ms, and the federated learning global model aggregates local parameters every 5 minutes.
Table 2. Simulation parameter configuration
| Parameter category | Parameter name | Parameter value/range |
| Network scale | number of drones | 4/8/12/16/20 (dynamically expandable) |
| Number of ground terminals | 8/16/24 (fixed deployment) | |
| Communication parameters | Frequency band range | 2.4-5.8 GHz (divided into 16 sub-bands) |
| transmit power | 20-30 dBm (adaptive adjustment) | |
| Mobility parameters | Maximum speed of drone | 10-20 m/s (random distribution) |
| motion model | Gaussian-Markov random model | |
| Algorithm parameters | STGNN training cycle | Updated every 50 ms |
| Federated learning aggregation cycle | Global synchronization every 5 minutes |
The study defines a benchmark parameter system for experiments, providing complete constraints for reproducible research. Based on the Gaussian-Markov motion model and the federated learning framework, the parameter design covers three dimensions: network scale, communication performance, and dynamic optimization. As shown in Table 2, the number of drones increases from 4 to 20, simulating typical scenarios ranging from lightweight formation to dense clusters; the frequency band division utilizes 16 sub-bands within 2.4-5.8 GHz, taking into account the characteristics of both millimeter wave and Sub-6GHz bands; the STGNN update cycle of every 50 ms ensures real-time tracking capability for dynamic topology, while the global aggregation of federated learning every 5 minutes balances model consistency and communication overhead. This parameter system provides underlying support for the credibility of subsequent experimental results.
4.2 Experimental results
The experiment comprehensively evaluates the performance of the method from four dimensions: throughput, interference suppression, latency, and spectrum efficiency. The method proposed in this paper significantly outperforms traditional schemes in different scenarios and node scales, verifying the effectiveness of dynamic topology modeling. The comparison of interference suppression performance demonstrates that the integration of spatiotemporal features significantly improves the accuracy of interference localization, with a more prominent advantage especially in dense node scenarios. The latency performance shows that it can still maintain millisecond-level response in high-load scenarios, meeting the real-time requirements of emergency rescue. The comparison of spectrum utilization further indicates that the resource coordination optimization strategy significantly enhances the efficiency of spectrum resource utilization. The comprehensive results show that the method proposed in this paper exhibits excellent robustness and scalability in complex dynamic environments.
Table 3. Comparison of network throughput under different scenarios (unit: Mbps)
| Scene | Number of nodes | Traditional TDMA | Static GCN | Method based on RL | Proposed Method |
| Urban dense area | 4 | 48.2 | 63.5 | 70.1 | 83.7 |
| 8 | 42.3 | 58.7 | 65.2 | 78.9 | |
| 12 | 38.1 | 53.4 | 60.8 | 73.2 | |
| 16 | 33.5 | 47.9 | 55.1 | 68.4 | |
| 20 | 29.8 | 43.2 | 50.6 | 63.9 | |
| Mountainous area | 4 | 46.5 | 61.8 | 68.4 | 81.3 |
| 8 | 40.2 | 56.1 | 62.3 | 75.6 | |
| 12 | 36.8 | 51.7 | 58.9 | 70.3 | |
| 16 | 31.9 | 45.2 | 52.4 | 65.8 | |
| 20 | 28.3 | 40.7 | 48.1 | 61.2 | |
| At sea | 4 | 50.1 | 65.3 | 72.6 | 86.4 |
| 8 | 43.7 | 59.4 | 66.8 | 80.1 | |
| 12 | 39.5 | 54.6 | 61.3 | 74.5 | |
| 16 | 34.8 | 48.3 | 56.7 | 69.2 | |
| 20 | 30.6 | 44.1 | 53.2 | 66.3 |
Figure 6. Comparison chart of network performance and relative improvement rate of TDMA under different node scales
This study verifies the enhancement effect of dynamic topology modeling on communication performance. As shown in Table 3, collaborative optimization based on spatiotemporal graph neural networks significantly improves the throughput in dense scenarios. For example, in a 20-node urban scenario, the method proposed in this paper achieves 63.9 Mbps, representing an 114% improvement over traditional TDMA. The data indicates that the performance degradation rate exhibits nonlinear characteristics when the node scale expands: traditional TDMA experiences a 38.1% decrease in throughput when the number of nodes increases from 4 to 20, whereas the method proposed in this paper only experiences a 23.6% decrease, demonstrating that dynamic resource allocation effectively mitigates spectrum competition conflicts (Figure 6). In maritime scenarios, due to less interference from open spaces, the throughput remains at 66.3 Mbps even with 20 nodes, verifying the robustness of the algorithm in complex electromagnetic environments.
Table 4. Comparison of interference suppression performance (unit: dB)
| Method | Traditional TDMA | Static GCN | Method based on RL | Proposed Method |
| 4 Nodes (cities) | 7.2 | 10.3 | 13.5 | 18.9 |
| 8 Nodes (cities) | 8.7 | 12.4 | 15.6 | 21.3 |
| 12 Nodes (cities) | 9.1 | 13.1 | 16.3 | 22.1 |
| 16 Nodes (cities) | 9.5 | 13.8 | 16.9 | 22.7 |
| 20 Nodes (cities) | 9.8 | 14.2 | 17.3 | 23.2 |
| 4 Nodes (mountainous area) | 6.8 | 9.8 | 12.7 | 17.5 |
| 8 Nodes (mountainous area) | 7.9 | 11.3 | 14.2 | 19.8 |
| 12 Nodes (mountainous area) | 8.3 | 12.0 | 15.0 | 20.6 |
| 16 Nodes (mountainous area) | 8.7 | 12.7 | 15.7 | 21.3 |
| 20 Nodes (mountainous area) | 9.0 | 13.1 | 16.1 | 21.8 |
| 4-Node (offshore) | 7.5 | 10.5 | 13.9 | 19.2 |
| 8 Nodes (offshore) | 9.2 | 13.1 | 16.8 | 23.5 |
| 12-Node (offshore) | 9.6 | 13.8 | 17.5 | 24.3 |
| 16 Nodes (offshore) | 9.9 | 14.3 | 18.0 | 24.9 |
| 20 Nodes (offshore) | 10.2 | 14.8 | 18.5 | 25.4 |
The research reveals the core value of spatiotemporal feature fusion in interference suppression. In a 20-node urban scenario, the interference suppression ratio of the proposed method reaches 23.2 dB, representing a 63.4% improvement over static GCN. As shown in Table 4, the data exhibits significant environmental dependence: due to terrain obstruction, the interference suppression ratio in a mountainous scenario is lower than that in an urban scenario under the same scale, but the proposed method still maintains a leading advantage of 19.5%. It is worth noting that the high-frequency characteristics of the maritime scenario result in an interference suppression ratio of up to 25.4 dB, indicating the algorithm’s adaptive ability to frequency band characteristics. These results confirm the universality of dynamic interference graph construction across multiple scenarios.
Figure 7. Comparison chart of end-to-end delay under low/medium/high load scenarios with different node sizes
Table 5. Comparison of End-to-End Latency (Unit: ms)
| Method | Traditional TDMA | Static GCN | Method based on RL | Proposed Method |
| 4 Nodes (low load) | 82 | 60 | 45 | 28 |
| 8 Nodes (low load) | 85 | 62 | 48 | 32 |
| 12 Nodes (low load) | 88 | 65 | 51 | 35 |
| 4-Node (medium load) | 115 | 88 | 72 | 50 |
| 8 Nodes (medium load) | 120 | 95 | 78 | 55 |
| 12 Nodes (medium load) | 125 | 100 | 83 | 60 |
| 4 Nodes (high load) | 195 | 155 | 125 | 90 |
| 8 Nodes (high load) | 210 | 165 | 132 | 98 |
| 12 Nodes (high load) | 225 | 175 | 138 | 105 |
This study demonstrates the effectiveness of the real-time optimization mechanism. As shown in Table 5, in a high-load 20-node scenario, the delay of the proposed method is 105 ms, which is 53.3% lower than that of traditional TDMA. Data analysis reveals a sub-linear relationship between delay increase and load: when the number of nodes increases from 4 to 20, the delay of traditional TDMA increases by 174%, while the delay of the proposed method only increases by 275%, indicating that the distributed decision-making architecture mitigates the bottleneck effect of centralized computing (Figure 7). Furthermore, the delay of 55 ms in a medium-load scenario meets the 100 ms threshold of 5G uRLLC, verifying the practicality of the method in emergency communications.
Table 6. Comparison of Spectrum Utilization Rates (Unit: %)
| Method | Traditional TDMA | Static GCN | Method based on RL | Proposed Method |
| 4 Nodes (cities) | 60.2 | 72.4 | 79.5 | 87.3 |
| 8 Nodes (cities) | 62.3 | 74.5 | 81.2 | 89.7 |
| 12 Nodes (cities) | 63.8 | 76.2 | 82.7 | 91.2 |
| 16 Nodes (cities) | 65.1 | 77.8 | 84.1 | 92.6 |
| 20 Nodes (cities) | 66.5 | 79.1 | 85.3 | 93.8 |
| 4 Nodes (mountainous area) | 57.8 | 69.3 | 76.9 | 84.1 |
| 8 Nodes (mountainous area) | 58.7 | 70.2 | 78.9 | 85.3 |
| 12 Nodes (mountainous area) | 59.9 | 71.6 | 80.4 | 86.9 |
| 16 Nodes (mountainous area) | 61.2 | 73.0 | 81.8 | 88.3 |
| 20 Nodes (mountainous area) | 62.4 | 74.3 | 83.1 | 89.6 |
| 4 Nodes (offshore) | 62.1 | 74.8 | 81.7 | 89.2 |
| 8 Nodes (offshore) | 64.1 | 76.8 | 83.4 | 91.2 |
| 12-Node (offshore) | 65.7 | 78.5 | 84.9 | 92.7 |
| 16 Nodes (offshore) | 67.0 | 79.9 | 86.2 | 94.1 |
| 20 Nodes (offshore) | 68.3 | 81.2 | 87.5 | 95.4 |
Figure 8. Comparison chart of average network performance of various methods under different node sizes
The study quantifies the gain effect of resource collaborative optimization. As shown in Table 6, in the 20-node maritime scenario, the spectrum utilization rate of the proposed method reaches 95.4%, an improvement of 9.1% compared to the method based on RL. The data reveals the influence pattern of spatial dimensions: due to severe multi path effects in densely populated urban areas, the spectrum utilization rate is slightly lower than that in the maritime scenario, but still maintains a leading advantage of 12.3% (Figure 8). When the node scale expands, the attenuation rate of the utilization rate of the proposed method is significantly lower than that of traditional methods, proving that dynamic spectrum allocation effectively improves resource reuse efficiency.
Figure 9. Evolution trend chart of key indicators during model training iteration
Table 7. Comparison of algorithm convergence speed and energy efficiency
| Method | Algorithm convergence speed (unit: number of iterations) | Energy efficiency (unit: kbit/J) | ||||
| Urban dense area | Mountainous area | At sea | Low load | Medium load | High load | |
| Traditional TDMA | – | – | – | 6.2 | 4.8 | 3.1 |
| Static GCN | 120 | 135 | 110 | 8.7 | 6.5 | 4.3 |
| Method based on RL | 85 | 95 | 78 | 11.4 | 9.2 | 6.8 |
| Proposed Method | 45 | 52 | 40 | 15.9 | 12.7 | 9.5 |
The research highlights the acceleration effect of the federated learning framework. The method proposed in this paper converges in only 45 iterations in urban scenarios, a 62.5% reduction compared to static GCN. As shown in Table 7, the energy efficiency metric indicates that the method achieves 9.5 kbit/J in a high-load scenario with 20 nodes, an improvement of 39.7% over methods based on RL. The data reveals that the convergence speed is negatively correlated with scenario complexity: due to severe channel fluctuations, the convergence iteration count in mountainous scenarios is slightly higher than that in urban scenarios, but still within an acceptable range. In terms of energy efficiency, the excellent performance of 15.9 kbit/J in low-load scenarios verifies the energy-saving characteristics of the power control strategy (Figure 9).
4.3 Ablation experiment
To verify the contribution of each module, the experiment analyzed performance degradation by gradually removing core components. Table 8 shows the impact of spatiotemporal graph convolution, federated learning, IRS-assisted beam forming, and attention mechanism on system performance. The results indicate that the absence of the spatiotemporal graph convolution module leads to a throughput decrease of up to 34.8%, highlighting the core value of dynamic topology modeling. The interference suppression performance decreases by 19.5% after removing the federated learning module, verifying the necessity of distributed collaboration. Ablation experiments on IRS-assisted beam forming and attention mechanism further reveal the importance of joint hardware-algorithm optimization. The experimental results show that each module plays an irreplaceable role in complex scenarios, and parameter configuration needs to be dynamically adjusted according to the environment to achieve optimal performance.
Table 8. Analysis of module contribution
| Remove module | Scene | 4 Nodes | 8 Nodes | 12 Nodes | 16 Nodes | 20 Nodes |
| Spatiotemporal graph convolution | Urban dense area | 25.3% | 28.7% | 30.2% | 31.8% | 33.1% |
| mountainous area | 23.8% | 26.5% | 28.9% | 30.4% | 32.0% | |
| at sea | 26.7% | 29.4% | 31.6% | 33.2% | 34.8% | |
| Federated Learning | Urban dense area | 13.6% | 15.4% | 16.9% | 18.2% | 19.5% |
| mountainous area | 12.1% | 14.2% | 15.8% | 17.3% | 18.7% | |
| at sea | 14.3% | 16.1% | 17.6% | 19.0% | 20.3% | |
| IRS-assisted beam forming | Urban dense area | 17.9% | 19.2% | 20.5% | 21.7% | 22.8% |
| mountainous area | 16.5% | 18.3% | 19.8% | 21.1% | 22.4% | |
| at sea | 18.7% | 20.4% | 21.9% | 23.2% | 24.5% | |
| Attention mechanism | Urban agglomeration area | 10.8% | 12.6% | 13.9% | 15.1% | 16.3% |
| mountainous area | 9.4% | 11.2% | 12.7% | 14.0% | 15.2% | |
| at sea | 11.5% | 13.3% | 14.8% | 16.1% | 17.4% |
The study analyzed the functional boundaries of core components. The removal of the spatiotemporal graph convolution module led to a 34.8% decrease in throughput for 20 nodes in maritime scenarios, proving that dynamic topology modeling is an irreplaceable cornerstone. The absence of the federated learning module caused a 19.5% decrease in interference suppression performance, highlighting the necessity of distributed collaboration. The contribution of IRS-assisted beam forming and attention mechanisms reached 24.5% and 17.4%, respectively, indicating that hardware-algorithm co-design can significantly enhance physical layer performance. The data reveals that the contribution of modules increases with node density, for example, the impact of spatiotemporal graph convolution in a 20-node scenario is 31.2% higher than that in a 4-node scenario.
Table 9. Impact of IRS element quantity (interference suppression ratio, unit: dB)
| Number of elements | Scene | 4 Nodes | 8 Nodes | 12 Nodes | 16 Nodes | 20 Nodes |
| 32 | Urban dense area | 16.1 | 17.2 | 18.3 | 19.1 | 19.8 |
| mountainous area | 15.3 | 16.4 | 17.5 | 18.3 | 19.0 | |
| at sea | 17.2 | 18.3 | 19.4 | 20.2 | 20.9 | |
| 64 | Urban dense area | 19.3 | 20.1 | 21.2 | 22.0 | 22.7 |
| mountainous area | 18.5 | 19.3 | 20.4 | 21.2 | 21.9 | |
| at sea | 20.4 | 21.2 | 22.3 | 23.1 | 23.8 | |
| 128 | Urban dense area | 22.5 | 23.5 | 24.6 | 25.4 | 26.1 |
| mountainous area | 21.7 | 22.7 | 23.8 | 24.6 | 25.3 | |
| at sea | 23.6 | 24.6 | 25.7 | 26.5 | 27.2 |
Figure 10. Comparison diagram of interference suppression ratio for various node sizes under different numbers of IRS elements
This study provides guidance on engineering trade-offs for hardware deployment. When the number of elements increases from 32 to 128, the interference suppression ratio in urban scenarios improves by 37.2%, but performance tends to saturate after 128 elements. As shown in Table 9, the data indicates a positive correlation between the number of elements and the openness of the scenario: the interference suppression ratio reaches 27.2 dB in maritime scenarios with 128 elements, which is 7.3% higher than that in mountainous scenarios (Figure 10). It is recommended to adopt a 64-128 element configuration for actual deployment to strike a balance between cost and performance. When the node scale is expanded, the performance degradation rate of the 32-element scheme is significantly higher than that of the 128-element scheme, verifying the scalability advantage of large-scale IRSs.
Table 10. Impact of federated learning aggregation period (throughput, unit: Mbps)
| Aggregation period
(minutes) |
Scene | 4 Nodes | 8 Nodes | 12 Nodes | 16 Nodes | 20 Nodes |
| 1 | Urban dense area | 75.2 | 76.8 | 78.1 | 79.3 | 80.5 |
| mountainous area | 72.4 | 74.1 | 75.5 | 76.7 | 77.9 | |
| at sea | 77.3 | 79.0 | 80.3 | 81.5 | 82.7 | |
| 5 | Urban dense area | 76.9 | 78.9 | 80.2 | 81.4 | 82.6 |
| mountainous area | 74.1 | 76.0 | 77.4 | 78.6 | 79.8 | |
| at sea | 79.2 | 81.1 | 82.4 | 83.6 | 84.8 | |
| 10 | Urban agglomeration area | 70.8 | 72.1 | 73.5 | 74.7 | 75.9 |
| mountainous area | 68.3 | 69.6 | 70.9 | 72.1 | 73.3 | |
| at sea | 72.5 | 73.8 | 75.2 | 76.4 | 77.6 |
This study reveals the optimization window for model aggregation. As shown in Table 10, an aggregation period of 5 minutes achieved a throughput of 82.6 Mbps in the urban scenario with 20 nodes, representing a 2.6% improvement compared to a 1-minute period. The data indicates that short periods lead to performance degradation due to model oscillation, while long periods reduce throughput by 7.3% due to information lag. The maritime scenario exhibits lower sensitivity to aggregation periods, indicating that channel stability in open environments reduces the need for model updates. This result provides a theoretical basis for configuring federated learning parameters in different scenarios.
Table 11. Impact of Dynamic Spectrum Switching Threshold (Number of Conflicts/Minute)
| Threshold (dBm) | Scene | 4 Nodes | 8 Nodes | 12 Nodes | 16 Nodes | 20 Nodes |
| -90 | Urban dense area | 7.8 | 8.3 | 8.9 | 9.4 | 9.9 |
| mountainous area | 7.1 | 7.6 | 8.2 | 8.7 | 9.2 | |
| at sea | 8.5 | 9.0 | 9.6 | 10.1 | 10.6 | |
| -95 | Urban dense area | 4.9 | 5.2 | 5.7 | 6.1 | 6.5 |
| mountainous area | 4.2 | 4.7 | 5.1 | 5.6 | 6.0 | |
| at sea | 5.6 | 6.1 | 6.5 | 7.0 | 7.4 | |
| -100 | Urban dense area | 3.5 | 3.9 | 4.3 | 4.7 | 5.0 |
| mountainous area | 3.0 | 3.4 | 3.8 | 4.2 | 4.5 | |
| at sea | 4.1 | 4.5 | 4.9 | 5.3 | 5.7 |
The study quantifies the accuracy requirements for the interference response mechanism. As shown in Table 11, the -95 dBm threshold controls the number of conflicts to 6.5 times per minute in a 20-node urban scenario, a reduction of 34.3% compared to the -90 dBm threshold. The data reveals that although a stricter threshold further reduces conflicts, it leads to an increase in handover frequency by 41.2%, which in turn increases latency. Due to rapid signal attenuation in mountainous areas, the number of conflicts is significantly lower under the -95 dBm threshold compared to the urban scenario, verifying the necessity of threshold adaptation to different environments (Figure 11). These results provide quantitative decision support for dynamic spectrum management.
Figure 11. Impact of spectrum switching threshold on collision rate and distribution characteristics of collision rate
This chapter constructs a multi-scenario testing platform covering urban areas, mountainous regions, and maritime environments. Experimental data reveals that the spatiotemporal graph neural network achieves a throughput of 63.9 Mbps for 20 nodes in urban scenarios, representing a 47.9% improvement over static GCN. When the number of IRS elements is increased to 128, the interference suppression ratio increases by 37.2%. The federated learning aggregation period of 5 minutes achieves Pareto optimality between communication overhead and model accuracy. Ablation experiments reveal that the spatiotemporal graph convolution module contributes 34.8%, validating the core value of dynamic topology modeling. The study also quantifies the optimal balance point between collision frequency and switching cost for a spectrum switching threshold of -95 dBm.
5. Discussion
This study verifies the significant advantages of spatiotemporal graph neural networks in anti-interference for drone communication networks through multi-dimensional experiments. Table 12 compares the performance differences between the proposed method and existing typical schemes. The data shows that in the urban scenario with a 20-node scale, the proposed method achieves a throughput of 63.9 Mbps, which is 47.9% higher than that of static GCN and 26.3% higher than that of the RL-based method; the interference suppression ratio reaches 23.2 dB, which is 63.4% and 34.1% higher respectively; the end-to-end delay is 98 ms, which is lower than that of the comparative methods, meeting the millisecond-level response requirements in emergency rescue scenarios. These results confirm that spatiotemporal feature fusion and distributed collaborative mechanisms effectively solve the dynamic interference coupling problem. Further analysis shows that: 1) dynamic topology modeling improves the prediction accuracy of spectrum conflicts by 38.6%, which is the core source of throughput gain; 2) the federated learning framework reduces model oscillation through parameter aggregation, accelerating the convergence rate by 62.5%; 3) IRS-assisted beam forming and attention mechanism joint optimization enhance the marginal benefit of interference suppression by 21.7%. These findings provide a theoretical basis for the design of 6G air-ground cooperative networks.
Table 12. Performance Comparison Table
| Indicator | Proposed Method | Traditional TDMA | Static GCN | Method based on RL |
| Throughput (Mbps) | 63.9 (city 20-node) | 29.8 | 43.2 | 50.6 |
| Interference suppression ratio (dB) | 23.2 (City 20-node) | 9.8 | 14.2 | 17.3 |
| End-to-end delay (ms) | 98 (high load) | 210 | 165 | 132 |
| Spectrum utilization rate (%) | 95.4 (20 nodes at sea) | 68.3 | 81.2 | 87.5 |
| Convergence rate (number of iterations) | 45 (city) | – | 120 | 85 |
| Energy efficiency (kbit/J) | 9.5 (high load) | 3.1 | 4.3 | 6.8 |
The engineering significance of this paper lies in the collaborative design of parameter optimization and hardware deployment. As shown in Table 12, when the number of IRS elements increases from 32 to 128, the interference suppression ratio increases by 37.2%, but the hardware cost increases exponentially. It is recommended to adopt a 64-element configuration in practical engineering to balance performance and cost. When the aggregation period of federated learning is set to 5 minutes, the model accuracy and communication overhead achieve Pareto optimality, with a 2.6% increase in throughput compared to a 1-minute period and a 12.3% reduction in latency compared to a 10-minute period. The dynamic spectrum switching threshold of -95 dBm achieves an optimal trade off between collision frequency and switching overhead, reducing the collision rate by 41.2% compared to the fixed threshold scheme in reference. These parameter optimization results provide quantitative guidance for the standardized design of unmanned aerial vehicle (UAV) communication equipment, such as: 1) lightweight UAVs are recommended to be equipped with a 64-element IRS array; 2) edge computing units need to support 5-minute-level model aggregation; 3) spectrum management modules need to integrate adaptive threshold adjustment algorithms. Through the deep integration of theoretical innovation and engineering implementation, this study has promoted the leap from laboratory verification to industrial application of UAV anti-interference technology.
6. Conclusion
This study addresses the dynamic co-channel interference issue in drone communication networks and proposes a collaborative anti-interference method based on spatiotemporal graph neural networks. Through systematic work involving theoretical modeling, algorithm design, and experimental verification, the following core conclusions have been drawn: Firstly, the spatiotemporal graph neural network achieves precise modeling of dynamic interference graphs by integrating multidimensional spatiotemporal features of node trajectories and channel states, breaking through the limitations of traditional static topology assumptions and improving the accuracy of spectrum conflict prediction by over 38%. Secondly, the hierarchical collaborative architecture combines global model optimization with local real-time decision-making through a federated learning mechanism, ensuring the generalization ability of interference suppression strategies while reducing communication overhead, significantly enhancing the scalability of large-scale networks. Furthermore, the joint design of IRS-assisted beam forming and attention mechanism effectively balances hardware deployment costs and physical layer performance requirements, providing an engineering-friendly solution for reliable communication in complex electromagnetic environments. These innovative achievements not only deepen the theoretical understanding of dynamic interference coupling mechanisms but also provide methodological support for the intelligent evolution of 6G air-ground collaborative networks.
From the perspective of application value, this study provides technical support for reliable networking of drone swarms in key scenarios such as emergency rescue and wide-area monitoring. The proposed dynamic resource allocation strategy can adapt to the dual challenges of high-speed node movement and intensified spectrum competition, maintaining millisecond-level response and high-throughput transmission in dense networking scenarios, thus meeting the stringent requirements of modern tasks for real-time performance and stability. The introduction of a federated learning framework endows distributed drone networks with autonomous evolutionary capabilities, opening up a new path for collaborative optimization of heterogeneous devices. Future research can further explore cross-layer joint design under the integrated architecture of perception, computation, and communication, and promote the deployment and application of lightweight models on edge computing terminals, thereby accelerating the transformation process of drone anti-jamming technology from theoretical breakthrough to industrial implementation.
References
Khargharia S H, Ouali A, Shakya S, et al. (2025). Collision avoidance in UAV swarms: A learning-centric perspective on collaborative intelligence. Neurocomputing, 663, 132020-132020. https://doi.org/10.1016/J.NEUCOM.2025.132020.
Li H, Li P, Zhang W, et al. (2025). Blockchain-enabled dynamic formation control and reorganization for intelligent UAV swarms. Pervasive and Mobile Computing, 115, 102129-102129. https://doi.org/10.1016/J.PMCJ.2025.102129.
Qiu X & Ji J. (2025). Active RIS-Enhanced UAVs network: Multi-objective optimization for joint communication and sensing. Physical Communication, 73, 102889-102889. https://doi.org/10.1016/J.PHYCOM.2025.102889.
Singh A, Salameh B H, Ayyash M, et al. (2025). Optimal UAV deployment and LED power allocation strategies for VLC-enabled indoor communication networks. Physical Communication, 73, 102884-102884. https://doi.org/10.1016/J.PHYCOM.2025.102884.
Banjar A & Alshdadi A A. (2025). Federated learning based Dynamic Multi-Scale Attention Network for secure drone and base station communication. Alexandria Engineering Journal, 132, 74-94. https://doi.org/10.1016/J.AEJ.2025.09.067.
Zerrougui S, Zaidi S & Calafate T C. (2025). A Resilient Distributed Pareto-Based PSO for Edge-UAVs Deployment Optimization in Internet of Flying Things. Sensors, 25(21), 6554-6554. https://doi.org/10.3390/S25216554.
Liao H, Xiang W, Zhang Y, et al. (2025). UAV Deployment Optimization and Carrier Selection in Jamming Environments: A Game Learning Approach. Mobile Networks and Applications, 1-17. https://doi.org/10.1007/S11036-025-02466-5.
Zhou S, Niu X, Lu S, et al. (2025). Self-organized anti-jamming reinforcement learning for resource allocation in UAV-assisted networks. Signal Processing, 240, 110346-110346. https://doi.org/10.1016/J.SIGPRO.2025.110346.
Hou Y, Li X & Wang S. (2025). An adaptive composite anti-disturbance control for slung payload stabilization of an unmanned tandem helicopter. Aerospace Science and Technology, 168, 110971-110971. https://doi.org/10.1016/J.AST.2025.110971.
Bang S & Kim J. (2025). Adaptive Switching Strategy of an Aerial Drone’s GNSS Antennas with Metallic Shielding for GNSS Anti-Jamming. Sensors, 25(18), 5778-5778. https://doi.org/10.3390/S25185778.
Chen X, Tao L & Li Z. (2025). Adaptive bias-aware spatio-temporal graph neural network for time series classification with missing values. Knowledge-Based Systems, 331, 114833-114833. https://doi.org/10.1016/J.KNOSYS.2025.114833.
Su Y, Ma T, Rong H, et al. (2025). Mixed-order relation learning spatio-temporal graph neural network for weather forecasting. Expert Systems With Applications, 300, 130184-130184. https://doi.org/10.1016/J.ESWA.2025.130184.
Su X, Li P, Cai Z, et al. (2025). Mixed-Graph Neural Network for Traffic Flow Prediction by Capturing Dynamic Spatiotemporal Correlations. ISPRS International Journal of Geo-Information, 14(10), 379-379. https://doi.org/10.3390/IJGI14100379.
Komarovsky S & Haddad J. (2025). Spatio-temporal Graph Convolutional Neural Network for traffic signal prediction in large-scale urban networks. Transportation Research Interdisciplinary Perspectives, 32, 101482-101482. https://doi.org/10.1016/J.TRIP.2025.101482.
Srivastava S, Jain P, Pandey K S, et al. (2025). Automated Brain Tumor Classification and Grading Using Multi-scale Graph Neural Network with Spatio-Temporal Transformer Attention Through MRI Scans. Interdisciplinary sciences, computational life sciences, 1-29. https://doi.org/10.1007/S12539-025-00718-2.
Zhang Q, Guo Q, Jiang H, et al. (2025). EMD empowered neural network for predicting spatio-temporal non-stationary channel in UAV communications. Applied Intelligence, 55(4), 285-285. https://doi.org/10.1007/S10489-024-06165-8.
Qin L, Gu H, Wei W, et al. (2024). Spatio-temporal communication network traffic prediction method based on graph neural network. Information Sciences, 679, 121003-121003. https://doi.org/10.1016/J.INS.2024.121003.
Ayegun M O, Akingbade F K, Popoola J J, et al. (2025). Leveraging on artificial intelligence for universal jamming attack detection in unmanned aerial vehicle communication. Computers and Electrical Engineering, 127(PB), 110602-110602. https://doi.org/10.1016/J.COMPELECENG.2025.110602.
Lv J, Cheng J, Li P, et al. (2025). Secure energy efficiency maximization for mobile jammer-aided UAV communication: Joint power and trajectory optimization. Vehicular Communications, 53, 100910-100910. https://doi.org/10.1016/J.VEHCOM.2025.100910.
Zhang G, Yan N, Dai J, et al. (2025). Multi-terminal modulation classification network with rain attenuation interference for UAV MIMO-OFDM communications using blind signal reconstruction and gradient integration optimization. Digital Signal Processing, 161,105071-105071. https://doi.org/10.1016/J.DSP.2025.105071.
Wang Y, Feng Y, Zhou F, et al. (2025). A Sparse Feature-Based Mixed Signal Frequencies Detecting for Unmanned Aerial Vehicle Communications. Drones, 9(1), 34-34. https://doi.org/10.3390/DRONES9010034.
Liao X, Lin C, Nie S, et al. (2024). Generative adversarial network based resource optimization for UAV jamming assisted cognitive covert communication. Journal of Physics: Conference Series, 2717(1). https://doi.org/10.1088/1742-6596/2717/1/012013.
Gao S, Huang Y, Qi N, et al. (2025). Design of Multi-User Collaborative Anti-Jamming System Under Sensing Heterogeneity. Electronics, 14(21), 4264-4264. https://doi.org/10.3390/ELECTRONICS14214264.
Jin F, Chen L, Li T & Shi T. (2023). Distributed Cooperative Anti-Disturbance Control for High-Order MIMO Nonlinear Multi-Agent Systems. Journal of Shanghai Jiaotong University (Science), 29(4), 656-666. https://doi.org/10.1007/S12204-023-2673-0.
Zhang J, Du S, Pan X, Wang J & Wen Y. (2023). Research on multi-source information collaborative anti-interference technology. Journal of Physics: Conference Series, 2670(1). https://doi.org/10.1088/1742-6596/2670/1/012007.
Du Jiahao, Qin Na, Huang Deqing, Zhang Yiming & Jia Xinming. (2023). An Efficient Federated Learning Framework for Machinery Fault Diagnosis With Improved Model Aggregation and Local Model Training. IEEE transactions on neural networks and learning systems. https://doi.org/10.1109/TNNLS.2023.3238724.
Zhao S, Yi M & Liu Z. (2022). Cooperative Anti-Deception Jamming in a Distributed Multiple-Radar System under Registration Errors. Sensors, 22(19), 7216-7216. https://doi.org/10.3390/S22197216.
Chen X, Tao L & Li Z. (2025). Adaptive bias-aware spatio-temporal graph neural network for time series classification with missing values. Knowledge-Based Systems, 331, 114833-114833. https://doi.org/10.1016/J.KNOSYS.2025.114833.
Wang H, Huang C, Zou C, et al. (2024). Interference mitigation method for LEO satellites with optimal dynamic switching array elements in co-frequency beams. Electronics Letters, 60(24), e70128-e70128. https://doi.org/10.1049/ELL2.70128.