Federated Learning for CSS

The foundational principle of Federated Learning for CSS is that raw sensing data never leaves the local node. Each cognitive radio computes a model update—typically gradient vectors—on its own observed spectrum data. Only these mathematical updates are transmitted to the fusion center, ensuring that sensitive intercepted waveforms, geolocation metadata, and proprietary signal characteristics remain on-device. This architecture is critical for defense and commercial operators who must comply with strict data sovereignty regulations while still benefiting from collaborative detection of low-signal-to-noise-ratio primary users.

The FedAvg algorithm serves as the standard fusion mechanism, replacing traditional hard or soft decision combiners. In each communication round:

• The fusion center distributes the current global model weights to selected sensing nodes.
• Each node performs local stochastic gradient descent on its private dataset.
• Nodes return their updated weights to the fusion center.
• The fusion center computes a weighted average of the local models, typically proportional to each node's dataset size, to produce a new global model. This iterative process converges to a classifier that generalizes across diverse radio environments without centralizing data.

A core challenge in Federated Learning for CSS is that local spectrum data is non-Independently and Identically Distributed (non-IID). Different sensing nodes experience distinct propagation environments, interference patterns, and primary user activity. One node may primarily observe AWGN noise floors while another encounters heavy multipath fading. Advanced algorithms like FedProx introduce a proximal term to the local objective function, preventing local models from diverging too far from the global model during training. This stabilizes convergence when data distributions are statistically heterogeneous across the cooperative network.

Even though raw data is not shared, model updates can inadvertently leak information through gradient inversion attacks. To provide formal privacy guarantees, Federated Learning for CSS integrates differential privacy (DP) mechanisms:

• Gradient clipping bounds the influence of any single training example.
• Gaussian noise is injected into the model updates before transmission.
• The privacy budget (ε, δ) quantifies the tradeoff between model accuracy and information leakage. This ensures that an adversary intercepting model updates cannot reliably infer whether a specific signal of interest was present in any node's local training set.

The reporting channel in CSS is often bandwidth-constrained, making the transmission of full model weights impractical. Federated Learning for CSS employs gradient compression techniques to reduce communication overhead:

• Gradient sparsification transmits only the top-k largest gradient values, zeroing out the rest.
• Quantization reduces each weight update from 32-bit floating point to 8-bit or even 1-bit representations.
• Periodic aggregation allows nodes to perform multiple local epochs between communication rounds, reducing the total number of uplink transmissions. These methods enable practical deployment over low-bandwidth control channels without sacrificing detection accuracy.

Federated Learning for CSS inherits vulnerability to Spectrum Sensing Data Falsification (SSDF) attacks, where malicious nodes submit corrupted model updates to poison the global classifier. Byzantine-resilient aggregation rules replace simple weighted averaging to mitigate this threat:

• Krum selects the update that is closest to a majority of its neighbors in vector space.
• Trimmed Mean discards the most extreme values for each model parameter before averaging.
• Median-based aggregation replaces the mean with the coordinate-wise median, which is inherently robust to outliers. These techniques ensure the global spectrum occupancy classifier remains reliable even when a fraction of cooperating nodes are compromised.

The foundational aggregation algorithm where a fusion center initializes a global model and distributes it to sensing nodes. Each node trains locally on its own IQ samples and returns only the updated model weights. The server then computes a weighted average of these updates to produce a new global model.

• Weighted by dataset size to prioritize nodes with more training data
• Multiple local epochs reduce communication rounds vs. standard SGD
• Non-IID robustness is critical since spectrum data distributions vary by node location

A mathematical framework that injects calibrated noise into model updates before transmission, providing formal guarantees that an adversary cannot infer whether a specific sensing node's data was included in training.

• Epsilon (ε) parameter controls the privacy-utility tradeoff; lower ε means stronger privacy
• Gaussian mechanism adds noise proportional to the sensitivity of the update
• Local DP applies noise at each node; global DP applies it at the aggregation server
• Prevents spectrum usage pattern leakage that could reveal operational tempos

Cryptographic techniques that ensure the fusion center can only compute the sum of model updates without inspecting any individual node's contribution in plaintext. This protects against honest-but-curious servers.

• Secret sharing splits each update into fragments distributed among peers
• Pairwise masking cancels out random masks during aggregation to reveal only the sum
• Dropout robustness handles nodes that go offline mid-protocol without stalling the round
• Critical for multi-tenant spectrum sharing where operators are mutual competitors

Spectrum data across distributed sensing nodes is inherently non-Independent and Identically Distributed due to varying propagation environments, hardware impairments, and primary user activity patterns. Federated learning for CSS must explicitly address this statistical heterogeneity.

• FedProx adds a proximal term to local objectives, limiting divergence from the global model
• SCAFFOLD uses control variates to correct for client drift during local training
• Clustered FL groups nodes with similar data distributions before aggregation
• Unaddressed non-IIDness causes catastrophic divergence in the global occupancy classifier

Transmitting full model weights over bandwidth-constrained reporting channels is prohibitive. Communication-efficient techniques compress or sparsify updates before transmission.

• Gradient quantization reduces weight precision to 8-bit or even 1-bit representations
• Gradient sparsification transmits only the top-k largest weight updates per round
• Periodic aggregation allows multiple local epochs between communication rounds
• Over-the-air computation exploits waveform superposition in wireless channels to compute the sum directly at the physical layer

Defense mechanisms against Spectrum Sensing Data Falsification (SSDF) attacks where malicious nodes submit poisoned model updates to corrupt the global occupancy classifier.

• Krum selects the update closest to a majority of its neighbors, ignoring outliers
• Trimmed Mean discards the most extreme values per parameter before averaging
• Median aggregation replaces the weighted mean with the coordinate-wise median
• Reputation scoring tracks historical update quality and down-weights suspect nodes
• These techniques mirror reputation management in traditional cooperative sensing but operate on gradient spaces rather than binary decisions