I present FLAP-Net, a federated learning framework specifically designed for automatic modulation recognition in China UAV swarm communication systems. As China UAV technology continues to advance rapidly, the demand for intelligent, real-time spectrum sensing capabilities in dense and contested electromagnetic environments grows proportionally. Automatic modulation recognition is a cornerstone of cognitive radio, enabling swift adaptation to dynamic channel conditions. However, conventional deep learning approaches are often too resource-intensive for the constrained edge platforms typical of China UAV nodes. To address this gap, I introduce FLAP-Net, which seamlessly integrates a lightweight classifier Feather-MSA, a cosine similarity-guided adaptive pruning mechanism, and a channel-aware weighted aggregation algorithm. Experimental results on the RadioML2016.10b dataset demonstrate that FLAP-Net achieves 93% recognition accuracy at a signal-to-noise ratio of 4 dB, with a single-sample inference latency of only 2.42 µs and a peak communication bandwidth below 27 Mb/s per China UAV. The framework exhibits exceptional robustness under non-IID data distributions and severe channel impairments, making it highly suitable for real-time collaborative spectrum awareness in China UAV swarms.
Introduction
China UAV swarms are increasingly deployed in electronic warfare, disaster relief, and cooperative communication missions. These dynamic networks demand rapid coordination, robust wireless links, and the ability to adapt to varying interference conditions. Cognitive radio techniques, such as automatic modulation recognition, allow receivers to identify unknown modulation schemes without prior knowledge of transmission parameters. Traditional likelihood-based methods offer strong theoretical performance but suffer from high computational complexity and sensitivity to model assumptions. Feature-based methods are lighter but degrade significantly under low SNR or multipath fading. Deep learning has emerged as a promising alternative, with convolutional networks achieving remarkable accuracy in open benchmarks. Nevertheless, the computational and memory demands of state-of-the-art models exceed the capabilities of China UAV edge devices. Federated learning provides a decentralized training paradigm that preserves data privacy while distributing computation across local nodes. However, applying FL to automatic modulation recognition presents two major challenges: the non-IID nature of data collected by different China UAVs, and the excessive communication overhead and model complexity. My proposed FLAP-Net addresses these issues through a set of carefully designed innovations.
System Model
Consider a China UAV swarm consisting of K aerial nodes communicating with a central base station. The received discrete-time baseband signal at the l-th UAV can be modeled as:
$$v[n] = \sum_{l=1}^{L} \alpha_l e^{j(2\pi f_{d,l} n + \phi_l)} s[n – d_l] + \eta[n]$$
where \alpha_l, d_l, f_{d,l}, and \phi_l are the channel gain, path delay, Doppler frequency, and phase offset for the l-th path, and \eta[n] is additive white Gaussian noise. The signal s[n] is modulated using one of several possible schemes from a set \mathcal{A}. The goal of automatic modulation recognition is to identify the correct scheme from the observed I/Q samples, which are represented as a 2 × 128 real-valued matrix:
$$X_{IQ} = \begin{bmatrix} X_I(0) & X_I(1) & \dots & X_I(N-1) \\ X_Q(0) & X_Q(1) & \dots & X_Q(N-1) \end{bmatrix}$$
The federated learning process proceeds as follows:
- The base station initializes global model weights \theta_{\text{global}}.
- Each China UAV receives the global weights and initializes its local Feather-MSA classifier.
- Local models are trained on distributed datasets, which are partitioned according to a Dirichlet distribution (parameter β) to simulate non-IID conditions.
- After local training, each China UAV applies adaptive pruning to reduce model size, then transmits the pruned weights \theta_{\text{local}}^{(k)} to the base station.
- The base station aggregates all pruned models using a channel-aware weighted aggregation algorithm to produce the updated global model.
Methodology and Optimization
Feather-MSA Local Classifier
I design Feather-MSA as a lightweight architecture tailored for China UAV platforms. The core of the classifier is a multi-scale convolution module with Squeeze-and-Excitation (SE) enhancement. Two parallel convolutional kernels with sizes 3×1 and 5×1 each produce 16 channels. After concatenation, batch normalization, and ReLU activation, an SE block adaptively recalibrates channel responses. A stride-2 convolution then performs temporal downsampling, reducing the tensor shape from (B, 2, T) to (B, 2, T/2). The initial feature representation F_0 is:
$$F_0 = f_{ds}(\sigma(\text{SE}(\text{BN}(F_{ms}))))$$
$$F_{ms} = \text{Concat}(\text{Conv}_{3\times1}(X), \text{Conv}_{5\times1}(X))$$
To capture temporal dependencies in China UAV communication signals, I incorporate a lightweight BiGRU module with 64 hidden units per direction. The hidden states are:
$$H = \text{BiGRU}(F_0)$$
A soft additive attention mechanism computes attention weights:
$$a_t = \text{softmax}(W_a h_t)$$
The final context vector is the weighted sum of hidden states:
$$\hat{F} = \sum_{t=1}^{T} a_t h_t$$
This attention mechanism allows the classifier to focus on the most informative time steps while maintaining a low computational footprint. Table 1 compares the performance of Feather-MSA against twelve baseline models on both RadioML2016.10a and RadioML2016.10b datasets.
| Model | Parameters | Acc a (%) | Acc b (%) | Loss a | Loss b | Time a (μs) | Time b (μs) |
|---|---|---|---|---|---|---|---|
| Feather-MSA | 43,643 | 62.27 | 64.59 | 1.0119 | 0.8595 | 2.13 | 2.42 |
| 1DCNN-PF | 174,794 | 58.17 | 53.45 | 1.0458 | 1.2547 | 41.25 | 33.74 |
| CGDNet | 124,676 | 59.68 | 54.73 | 0.9918 | 1.1876 | 102.04 | 22.92 |
| CLDNN | 163,976 | 58.47 | 55.84 | 1.0139 | 1.1698 | 110.24 | 26.82 |
| CNN1 | 1,592,126 | 58.68 | 55.68 | 0.9898 | 1.1733 | 11.17 | 9.91 |
| CNN2 | 857,994 | 59.72 | 57.54 | 0.9673 | 1.1167 | 56.06 | 55.89 |
| DenseNet | 3,282,474 | 58.84 | 54.50 | 1.0137 | 1.1960 | 143.55 | 144.90 |
| IC-AMCNET | 1,263,882 | 60.89 | 56.46 | 0.9624 | 1.1383 | 17.92 | 16.86 |
| MCNET | 121,226 | 60.03 | 54.36 | 0.9799 | 1.2289 | 19.04 | 15.56 |
| ResNet | 3,098,154 | 59.48 | 54.57 | 0.9998 | 1.2188 | 93.87 | 105.67 |
Feather-MSA achieves the highest accuracy with the fewest parameters and lowest inference latency, making it ideal for resource-constrained China UAV edge nodes.
Adaptive Pruning Mechanism
To minimize communication overhead in China UAV swarms, I introduce a cosine similarity-guided adaptive pruning method. For each local model with weights \theta_{\text{local}} and global model weights \theta_{\text{global}}, the cosine similarity is:
$$\cos_{\theta_{LG}} = \frac{\langle \theta_{\text{local}}, \theta_{\text{global}} \rangle}{\|\theta_{\text{local}}\| \|\theta_{\text{global}}\|}$$
The dynamic pruning ratio P is computed as:
$$P = \gamma \cdot (1 – \cos_{\theta_{LG}})$$
where \gamma controls the pruning intensity. The number of pruned parameters n_P is:
$$n_P = P \times \phi$$
where \phi is the total number of parameters. This mechanism allows each China UAV to selectively remove weights that deviate significantly from the global consensus, thereby reducing transmitted data while preserving essential knowledge.
Channel-Aware Weighted Aggregation
After receiving pruned local models, the base station computes a channel quality function q(\gamma_k) using a sigmoid mapping:
$$q(\gamma_k) = \frac{1}{1 + e^{-\alpha(\gamma_k – \delta)}}$$
where \gamma_k is the instantaneous SNR of the k-th China UAV, \alpha is a scaling factor, and \delta is a threshold. The global aggregation is then:
$$\theta_{\text{global}}^{(r+1)} = \frac{\sum_{k=1}^{K} n_k q(\gamma_k) \theta_{\text{local}}^{(k)}}{\sum_{j=1}^{K} n_j q(\gamma_j)}$$
where n_k is the number of local samples at China UAV k. This dynamic weighting ensures that nodes with higher SNR contribute more to the global model, mitigating the impact of noisy or misaligned updates.
End-to-End Algorithm
The complete FLAP-Net training process is summarized in Algorithm 1, which iterates over R communication rounds until convergence.
Algorithm 1: FLAP-Net Adaptive Pruning and Federated Learning
- Initialize global model \theta_{\text{global}} randomly.
- For each round r = 1 to R:
-
- Broadcast \theta_{\text{global}} to all China UAV nodes.
- For each node k in parallel:
-
- \theta_{\text{local}} \gets \theta_{\text{global}}
- Train Feather-MSA on local data D_k for E epochs to minimize cross-entropy loss.
- Apply adaptive pruning based on cosine similarity.
- Base station aggregates pruned models using channel-aware weighted aggregation.
- If \|\theta_{\text{global}}^{(r+1)} – \theta_{\text{global}}^{(r)}\| < \epsilon, break.
- Update \theta_{\text{global}}.
- Output optimized global model \theta_{\text{global}}^*.
Experimental Results and Analysis
Dataset and Simulation Setup
I evaluate FLAP-Net using the RadioML2016.10b dataset, which contains 1.2 million I/Q samples from 10 modulation types, each represented as a 2×128 matrix. SNR varies from -20 dB to +18 dB in 2 dB steps. I simulate a synchronous FL system with K = 5 China UAV nodes. Key parameters are listed in Table 2.
| Parameter | Value |
|---|---|
| Number of UAVs (K) | 5 |
| Communication rounds (R) | 15 |
| Local epochs (E) | 10 |
| Optimizer | AdamW |
| Learning rate | 0.001 |
| Batch size | 128 |
| Loss function | Cross-entropy |
| Data distribution | Dirichlet (β ∈ {1.0,0.5,0.3,0.1,0.01}) |
Feather-MSA Performance Verification
Feather-MSA achieves 64.59% overall accuracy on the full SNR range of RadioML2016.10b, outperforming all baselines. At 4 dB SNR, the accuracy reaches 93%. The per-modulation accuracy curves demonstrate consistent superiority across all schemes, especially for high-order modulations such as 16QAM and 64QAM. The inference latency of 2.42 µs per sample enables real-time operation on China UAV platforms.
End-to-End System Evaluation
Table 3 reports the communication bandwidth per China UAV node. The peak bandwidth stays below 27 Mb/s for each node, while the end-to-end average bandwidth over the entire training duration is only 0.66 Mb/s. This dramatic reduction is attributable to the adaptive pruning mechanism, which selectively transmits only the most relevant model parameters.
| UAV Node | Average BW (Mbps) | Peak BW (Mbps) |
|---|---|---|
| Node 0 | 26.70 | 26.70 |
| Node 1 | 26.70 | 26.70 |
| Node 2 | 26.69 | 26.69 |
| Node 3 | 26.70 | 26.70 |
| Node 4 | 26.69 | 26.69 |
| End-to-end average | 0.66 Mbps | |
Impact of Pruning Intensity
I vary the pruning intensity parameter γ from 0.0 (no pruning) to 1.0 (aggressive pruning). As shown in Table 4, accuracy remains above 64.57% across all γ values. The best performance occurs at γ = 0.3 and γ = 0.6, indicating that moderate pruning acts as a beneficial regularizer. Even at γ = 0.8, the accuracy degrades by less than 0.05%, confirming that the adaptive pruning mechanism effectively retains discriminative features.
| γ | Validation Acc (%) | Test Acc (%) |
|---|---|---|
| 0.0 | 64.58 | 64.57 |
| 0.3 | 64.62 | 64.62 |
| 0.6 | 64.59 | 64.59 |
| 0.8 | 64.58 | 64.57 |
| 1.0 | 64.57 | 64.57 |
Robustness to Non-IID Data
Using Dirichlet partitioning with varying β values, I evaluate FLAP-Net under increasingly heterogeneous data distributions. Table 5 summarizes the results. At β = 0.01 (extreme non-IID), the overall accuracy drops from 64.59% to 61.71%, but still exceeds 91% at SNR ≥ 0 dB. This resilience is attributed to the combination of lightweight architecture, adaptive pruning, and channel-aware aggregation, which collectively mitigate the effects of local data skew.
| β | Full SNR (%) | SNR ≥ 0 dB (%) |
|---|---|---|
| 1.0 | 64.59 | 93.42 |
| 0.5 | 64.35 | 92.87 |
| 0.3 | 64.30 | 92.41 |
| 0.1 | 62.57 | 91.85 |
| 0.01 | 61.71 | 91.23 |
Aggregation Strategy Comparison
I compare FLAP-Net’s channel-aware aggregation against FedAvg, FedProx, and FedBN under β = 0.3. Table 6 shows that FLAP-Net achieves 64.30% overall accuracy, outperforming FedAvg (57.38%), FedProx (61.02%), and FedBN (53.87%). The performance gap is particularly pronounced in low-to-mid SNR ranges, where noisy updates from poor-quality nodes are effectively suppressed by the channel-aware weights.
| Method | Full SNR (%) | SNR ≥ 0 dB (%) |
|---|---|---|
| FLAP-Net | 64.30 | 92.41 |
| FedAvg | 57.38 | 88.14 |
| FedProx | 61.02 | 90.23 |
| FedBN | 53.87 | 86.46 |
Conclusion
I have presented FLAP-Net, a lightweight and communication-efficient federated learning framework for automatic modulation recognition in China UAV swarm communications. The framework integrates a minimalistic yet powerful Feather-MSA classifier, a cosine similarity-guided adaptive pruning strategy, and a channel-aware weighted aggregation algorithm. Extensive experiments on radio frequency datasets demonstrate that FLAP-Net achieves superior recognition accuracy, ultra-low inference latency, and drastically reduced communication overhead compared to existing centralized and federated methods. The system remains robust under severe non-IID conditions and low-SNR scenarios, making it a viable solution for real-time collaborative spectrum awareness in China UAV networks. Future work will extend FLAP-Net to asynchronous federated learning, investigate its resilience against adversarial attacks, and validate its performance in field deployments of China UAV swarms.