In this work, I present a federated learning framework named FLAP-Net, specifically designed for automatic modulation recognition (AMR) in China drone swarm communication systems. The goal is to enable real-time, bandwidth-efficient, and accurate spectrum awareness for unmanned aerial vehicles operating in dynamic and contested environments. My approach integrates a lightweight local classifier called Feather-MSA, an adaptive pruning mechanism guided by cosine similarity, and a channel-aware weighted aggregation algorithm. Extensive experiments on the RadioML2016.10b dataset demonstrate that FLAP-Net achieves a classification accuracy of 93% at 4 dB signal-to-noise ratio (SNR) with an inference latency of only 2.42 µs per sample and a peak communication bandwidth below 27 Mb/s per China drone node. These results highlight the practicality of FLAP-Net for real-time collaborative AMR in China drone swarms.
China drone technology has advanced rapidly, and swarms of UAVs are increasingly deployed for intelligent wireless communication, electronic warfare, disaster rescue, and coordinated sensing. The ability to automatically recognize modulation schemes without prior knowledge is essential for cognitive radio systems in these swarms. However, deep learning-based AMR models are often too resource-intensive to run on the edge devices carried by China drones. Moreover, the non-independent and identically distributed (Non-IID) nature of data collected across different drones, combined with severe bandwidth constraints, poses significant challenges for traditional centralized or federated learning approaches. To address these issues, I propose FLAP-Net, which balances accuracy, efficiency, and communication overhead.
System Model
Consider a China drone swarm consisting of multiple UAV nodes communicating with a central base station (BS). The received baseband signal at each drone can be expressed as:
$$
v[n] = \sum_{l=1}^{L} \alpha_l e^{j(2\pi f_{d,l} n + \phi_l)} s[n – \tau_l] + \eta[n]
$$
where $\alpha_l$, $\tau_l$, $f_{d,l}$, and $\phi_l$ are the channel gain, path delay, Doppler frequency, and phase shift of the $l$-th path, respectively; $\eta[n]$ is additive white Gaussian noise; and $s[n]$ is the transmitted modulation signal. The cooperative reception across $K$ drones yields an enhanced signal representation:
$$
\mathbf{Y}(N) = \mathcal{H}(\mathbf{y}(N)) + \mathbf{n}(N)
$$
where $\mathbf{y}(N) = \{x_1(N), x_2(N), \dots, x_K(N)\}$ denotes the collection of received signals from all China drone nodes, and $\mathcal{H}(\cdot)$ represents a cooperative feature enhancement operation (e.g., beamforming or spatial diversity). For deep learning classification, the complex signal is decomposed into in-phase (I) and quadrature (Q) components:
$$
\mathbf{X}_{IQ} = \begin{bmatrix} X_I(0) & X_I(1) & \cdots & X_I(N-1) \\ X_Q(0) & X_Q(1) & \cdots & X_Q(N-1) \end{bmatrix}
$$
The task of AMR is to identify the modulation scheme $\mathcal{A}(\cdot)$ from the above I/Q representation. In the federated learning setting, each China drone node trains a local model on its own data and only shares model updates with the BS, preserving data privacy. The global model is updated via:
$$
\theta_{\text{global}}^{(r+1)} = \sum_{k=1}^{K} \lambda_k \cdot \theta_{\text{local}}^{(k,r)}
$$
where $\lambda_k$ are aggregation weights that I will design adaptively based on channel conditions.
Method and Optimization
Feather-MSA: Lightweight Local Classifier
I designed a lightweight classifier called Feather-MSA specifically for China drone edge devices. It comprises three key modules:
Multi-Scale Squeeze-and-Excitation (MS-SE) Block. This module uses two parallel convolutional kernels of sizes $3 \times 1$ and $5 \times 1$, each generating 16 channels. The feature maps are concatenated, batch-normalized, and passed through a Squeeze-and-Excitation block for adaptive channel recalibration. A stride-2 $3 \times 1$ convolution performs temporal downsampling, reducing the tensor from shape $(B, 2, T)$ to $(B, 2, T/2)$. The initial feature representation $\mathbf{F}_0$ is obtained as:
$$
\mathbf{F}_0 = f_{ds}(\sigma(\text{SE}(\text{BN}(\mathbf{X}_{\text{ms}}))))
$$
$$
\mathbf{X}_{\text{ms}} = \text{Concat}(\text{Conv}_{3\times1}(\mathbf{X}), \text{Conv}_{5\times1}(\mathbf{X}))
$$
where $\mathbf{X}$ is the input I/Q matrix, $\text{Conv}$ denotes convolution, $\text{BN}$ is batch normalization, $\text{SE}$ is the squeeze-and-excitation operation, $\sigma$ is ReLU, and $f_{ds}$ is the downsampling convolution.
Lightweight BiGRU Temporal Encoding Module. To capture temporal dependencies in China drone communication signals, I employ a single-layer bidirectional GRU with 64 hidden units per direction. The hidden states $\mathbf{H}$ are computed as:
$$
\mathbf{H} = \text{BiGRU}(\mathbf{F}_0)
$$
This design avoids the heavy computational cost of Transformer models while maintaining strong sequential modeling capability.
Additive Attention Mechanism. Instead of using fixed weights for the GRU outputs, I incorporate additive attention to focus on the most relevant time steps. The attention weight $a_t$ is:
$$
a_t = \text{softmax}(\mathbf{W}_a \mathbf{h}_t)
$$
The final context vector $\hat{\mathbf{F}}$ is the weighted sum of all hidden states:
$$
\hat{\mathbf{F}} = \sum_{t=1}^{T’} a_t \mathbf{h}_t
$$
This vector is then passed to a fully connected layer for modulation classification. Table 1 compares Feather-MSA with other models on RadioML2016.10b in terms of parameters, accuracy, and inference time.
Table 1: Performance comparison of different models on RadioML2016.10b dataset.
| Model | Parameters | Accuracy (%) | Inference Time (µs) |
|---|---|---|---|
| Feather-MSA | 43,643 | 64.59 | 2.42 |
| 1DCNN-PF | 174,794 | 53.45 | 33.74 |
| CGDNet | 124,676 | 54.73 | 22.92 |
| CLDNN | 163,976 | 55.84 | 26.82 |
| CLDNN2 | 517,314 | 56.12 | 51.96 |
| CNN1 | 1,592,126 | 55.68 | 9.91 |
| CNN2 | 857,994 | 57.54 | 55.89 |
| DAE | 14,812 | 53.68 | 33.21 |
| DenseNet | 3,282,474 | 54.50 | 144.90 |
| IC-AMCNET | 1,263,882 | 56.46 | 16.86 |
| MCNET | 121,226 | 54.36 | 15.56 |
| ResNet | 3,098,154 | 54.57 | 105.67 |
Feather-MSA achieves the highest accuracy with the fewest parameters (43,643) and the fastest inference (2.42 µs), making it ideal for resource-constrained China drone platforms.
Adaptive Pruning Mechanism
To reduce uplink communication overhead in China drone swarms, I propose an adaptive pruning mechanism that removes redundant model parameters based on cosine similarity between local and global weights. The cosine similarity is defined as:
$$
\cos\theta_{\text{LG}} = \frac{\langle \theta_{\text{local}}, \theta_{\text{global}} \rangle}{\|\theta_{\text{local}}\| \cdot \|\theta_{\text{global}}\|}
$$
A smaller $\cos\theta_{\text{LG}}$ indicates a larger divergence from the global consensus, suggesting that the local parameters contain more noise or are biased by Non-IID data. The pruning rate $P$ is dynamically computed as:
$$
P = \gamma \cdot (1 – \cos\theta_{\text{LG}})
$$
where $\gamma$ is a pruning strength factor. The number of weights to prune is $n_P = P \times \phi$, where $\phi$ is the total number of weights. This adaptive approach significantly reduces the number of parameters that need to be transmitted from each China drone to the BS, as summarized in Table 2.
Table 2: Bandwidth consumption per China drone node under different pruning strengths.
| Pruning Strength $\gamma$ | Average Bandwidth (Mbps) | Peak Bandwidth (Mbps) |
|---|---|---|
| 0.0 | 52.10 | 52.10 |
| 0.3 | 26.70 | 26.70 |
| 0.6 | 18.45 | 18.45 |
| 1.0 | 10.32 | 10.32 |
With $\gamma = 0.3$, the peak bandwidth is reduced from 52.10 Mbps to 26.70 Mbps, while maintaining accuracy within 0.03% of the unpruned model. This demonstrates the effectiveness of the adaptive pruning mechanism for bandwidth-limited China drone communication links.
Channel-Aware Weighted Aggregation
The BS aggregates local models from China drone nodes using a channel-aware weighting scheme. I define a channel quality function $q(\gamma_k)$ that maps the SNR of the $k$-th drone to a weight value using a sigmoid normalization:
$$
q(\gamma_k) = \frac{1}{1 + e^{-\alpha(\gamma_k – \delta)}}
$$
where $\alpha$ is a scaling factor and $\delta$ is a threshold. The global update rule is:
$$
\theta_{\text{global}}^{(r+1)} = \frac{\sum_{k=1}^{K} n_k \cdot q(\gamma_k) \cdot \theta_{\text{local}}^{(k,r)}}{\sum_{j=1}^{K} n_j \cdot q(\gamma_j)}
$$
Here $n_k$ is the number of local samples at drone $k$. This scheme reduces the contribution of China drones with poor channel quality (low SNR), mitigating the impact of noisy or corrupted updates. Table 3 compares the aggregation strategies in a Non-IID setting ($\beta = 0.3$).
Table 3: Average classification accuracy under different aggregation strategies.
| Aggregation Method | Full SNR Range (%) | SNR $\geq$ 0 dB (%) |
|---|---|---|
| FLAP-Net (proposed) | 64.30 | 92.41 |
| FedAvg | 57.38 | 88.14 |
| FedProx | 61.02 | 90.23 |
| FedBN | 53.87 | 86.46 |
FLAP-Net outperforms all baselines, especially in the critical high-SNR region where accurate aggregation matters most for China drone mission success.
End-to-End System Evaluation
I evaluated FLAP-Net in a simulated China drone swarm with $K=5$ nodes, using the RadioML2016.10b dataset. The training involved 15 communication rounds with 10 local epochs per round, using AdamW optimizer with initial learning rate 0.001 and batch size 128. The Non-IID data distribution was simulated using a Dirichlet distribution with concentration parameter $\beta$. I measured the accuracy across all 20 SNR levels from -20 dB to 18 dB.
The system achieves a peak accuracy of 93% at 4 dB SNR, and maintains over 90% accuracy for all SNR $\geq$ 0 dB. The per-class confusion matrix shows that lower-order modulations such as BPSK and QPSK are recognized with near-perfect accuracy at moderate SNR, while higher-order modulations like 64QAM and 256QAM also perform well. Table 4 summarizes the end-to-end bandwidth consumption.
Table 4: Communication bandwidth of FLAP-Net for each China drone node ( $\gamma = 0.3$).
| Node ID | Average Bandwidth (Mbps) | Peak Bandwidth (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 | |
The end-to-end average bandwidth (0.66 Mbps) is much lower than the instantaneous per-node bandwidth, because the training process includes computation-heavy local epochs that do not require transmission. This confirms that FLAP-Net is highly communication-efficient for China drone swarms.
Impact of Pruning Strength
I performed a sensitivity analysis on the pruning strength $\gamma$ from 0.0 to 1.0. Figure 8 (described in text) shows that the validation and test accuracy on RadioML2016.10b remain above 64.57% for all $\gamma$ values. Interestingly, moderate pruning ($\gamma = 0.3$ or $0.6$) yields slightly higher accuracy than no pruning, acting as a regularization effect. Even at aggressive pruning ($\gamma = 0.8$), the accuracy drop is less than 0.1%. This robustness is crucial for practical China drone deployments where channel conditions vary rapidly.
Robustness to Non-IID Data
I tested FLAP-Net under several Dirichlet parameters $\beta \in \{1.0, 0.5, 0.3, 0.1, 0.01\}$. Table 5 presents the average accuracy across all SNRs and for SNR $\geq$ 0 dB.
Table 5: Accuracy of FLAP-Net under different Dirichlet parameters $\beta$.
| $\beta$ | Full SNR Range (%) | SNR $\geq$ 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 |
Even under the most extreme Non-IID condition ($\beta = 0.01$), the accuracy for SNR $\geq$ 0 dB remains above 91%, demonstrating the robustness of FLAP-Net against data heterogeneity that is common in China drone swarm deployments.
Conclusion
In this work, I proposed FLAP-Net, a lightweight and communication-efficient federated learning framework for automatic modulation recognition in China drone swarm communication systems. The framework integrates a novel Feather-MSA local classifier that achieves state-of-the-art accuracy on the RadioML2016.10b dataset with only 43,643 parameters and 2.42 µs inference latency. The adaptive pruning mechanism based on cosine similarity effectively reduces uplink bandwidth to below 27 Mb/s per drone, and the channel-aware weighted aggregation algorithm improves global model accuracy under dynamic channel conditions. Extensive experiments confirm that FLAP-Net maintains high recognition accuracy even under severe Non-IID data distributions and at low SNRs. These features make FLAP-Net an ideal solution for real-time collaborative spectrum awareness in China drone swarms, paving the way for smarter and more resilient unmanned systems in future intelligent wireless networks.