Technical Summary: FLAME – A Federated Learning Approach for Multi-Modal RF Fingerprinting

1. Problem Statement

Radio Frequency (RF) fingerprinting is a critical technique for identifying transmitting devices based on unique hardware imperfections reflected in their transmitted waveforms. While deep learning has significantly outperformed traditional statistical methods in this domain, scaling these approaches in next-generation wireless networks presents two major challenges:

1. Privacy and Data Centralization: Aggregating massive datasets (gigabytes to terabytes) from multiple Access Points (APs) to a central server for training is computationally expensive and increases the risk of data leakage.
2. Federated Learning (FL) Limitations: While FL offers a privacy-preserving alternative by training models locally at APs, existing FL approaches for RF fingerprinting suffer from slow convergence and high communication overhead. This is largely because they rely on single-modal processing (typically In-phase/Quadrature or I/Q time samples). In non-independent and identically distributed (non-IID) settings, where APs observe different subsets of transmitters, single-modal FL requires excessive training rounds to achieve high accuracy, leading to inefficient training times and significant client drift.

Although multi-modal representations have shown promise in centralized RF fingerprinting, extending these approaches to FL is non-trivial due to data heterogeneity, constrained local data, and limited coordination between clients.

2. Methodology: The FLAME Framework

The authors propose FLAME (Federated Learning Approach for Multi-Modal RF Fingerprinting), a framework designed to integrate complementary RF signal representations within a federated learning paradigm.

Multi-Modal Signal Representation

Instead of relying solely on raw I/Q time samples, FLAME represents each received signal in three complementary modalities at every AP:

1. I/Q Time-Domain: The raw complex baseband signal mapped to a real-valued matrix (Real and Imaginary components).
2. Frequency-Domain: The Discrete Fourier Transform (DFT) of the signal, capturing spectral characteristics.
3. Amplitude-Phase (Phasor): The magnitude and phase of the signal at each time sample.

Architecture and Training

FLAME employs a late-fusion deep learning architecture:

• Modality-Specific Encoders: Each of the three modalities is processed by a dedicated encoder (e.g., a neural network branch).
• Shared Fusion and Decoder: The features from all encoders are concatenated and passed through a shared fusion module and a shared decoder to predict the transmitter identity.
• Federated Process:
  • Local models at each AP are trained using Stochastic Gradient Descent (SGD) on the concatenated multi-modal inputs.
  • Updated model parameters are transmitted to a central server.
  • The server aggregates parameters (e.g., via FedAvg) to update the global model, which is then redistributed to APs.

Theoretical Analysis

The paper provides a theoretical convergence analysis demonstrating that multi-modal FL can reduce the variance of the global loss.

• Cross-Modal Interaction: The authors define a "Multi-Modal Interaction Factor" ($\Gamma_{MM}$) that quantifies the alignment between local and global gradient directions.
• Convergence Bound: They derive a convergence bound showing that if modalities are complementary (inducing aligned gradient directions on shared parameters), $\Gamma_{MM}$ decreases. This leads to a tighter convergence bound, allowing for fewer communication rounds and faster convergence compared to single-modal FL. Conversely, if modalities are distractive (conflicting gradients), convergence may suffer.

3. Key Contributions

1. Framework Development: FLAME is presented as the first work to jointly integrate multi-modal RF signal representations (I/Q, DFT, and Amplitude-Phase) within a federated learning paradigm for RF fingerprinting.
2. Theoretical Feasibility: The authors theoretically demonstrate that incorporating multiple complementary modalities improves convergence rates and reduces loss variance compared to single-modal approaches. They establish a convergence bound that explicitly accounts for cross-modal interactions and data heterogeneity.
3. Empirical Validation: Extensive experiments on real-world datasets (ORBIT and ORACLE) validate the theoretical findings. The results show that FLAME consistently outperforms single-modal baselines and various modality combinations across different FL optimization algorithms (FedAvg, FedProx, SCAFFOLD) and data heterogeneity levels.

4. Experimental Results

The evaluation was conducted on two datasets: the ORBIT dataset (163 WiFi transmitters) and the ORACLE dataset (16 USRP transmitters).

• Performance Gains: FLAME achieved significant accuracy improvements over the best single-modal or partial multi-modal baselines. For instance, under extreme data heterogeneity ($\alpha=0.1$), FLAME showed a 9.3% improvement over the second-best modality combination. Under moderate heterogeneity ($\alpha=1$), the improvement was 6.8%.
• Robustness: The performance gains were consistent across different FL algorithms (FedAvg, FedProx, SCAFFOLD) and varying numbers of clients ($N=10, 20, 40$).
• Ablation Studies:
  • Capacity Matching: Even when the single-modal baseline was adjusted to have the same number of parameters (capacity-matched) as the multi-modal model, FLAME still outperformed it, confirming that the gains stem from complementary information rather than increased model size.
  • Local Rounds: FLAME maintained superior performance regardless of the number of local SGD iterations ($J=10$ vs $20$), whereas other methods showed sensitivity to local bias.
• Feature Space Analysis: t-SNE and PCA visualizations of the fused feature embeddings revealed that FLAME produces more distinct and well-separated class clusters compared to single-modal approaches, particularly in low-heterogeneity settings.
• Efficiency: While FLAME incurs higher local computation due to multiple encoders, it requires fewer communication rounds to reach specific accuracy thresholds compared to single-modal methods, effectively trading local compute for reduced communication overhead.

5. Significance and Claims

The paper claims that FLAME addresses the critical bottleneck of slow convergence in federated RF fingerprinting. By leveraging the complementary nature of different signal representations (time, frequency, and phase), the framework reduces the variance of the global loss, allowing local models to diverge less from the global model during training. This results in:

• Higher Accuracy: Improved device discrimination capabilities in distributed settings.
• Faster Convergence: Fewer training rounds and communication exchanges are required to achieve high performance.
• Privacy Preservation: The approach maintains the privacy benefits of FL while overcoming the performance limitations of single-modal data processing.

The authors acknowledge limitations, including the increased local computation cost and the need for future work on handling partial modalities (where devices cannot compute all representations) and robustness against adversarial attacks. However, they assert that FLAME provides a foundational step toward efficient, high-performance, and privacy-preserving RF fingerprinting in large-scale wireless networks.