Generative Adversarial Network (GAN) for RF

The core mechanism involves a minimax game between two neural networks. The Generator learns to map random noise to synthetic I/Q samples that mimic normal spectrum activity. The Discriminator learns to distinguish between real normal samples and the generator's fakes. This adversarial pressure forces the generator to capture the true data distribution, while the discriminator becomes an expert at identifying subtle deviations.

Once trained, the discriminator serves as a powerful anomaly detector. New spectrum captures are fed directly to the discriminator, which outputs an anomaly score representing the probability that the input is 'fake' or out-of-distribution. Key scoring methods include:

• Discriminator output: Direct probability of an input being synthetic.
• Feature matching: Comparing intermediate layer activations of real and test data.
• Reconstruction error: Using a bidirectional GAN (BiGAN) to map data back to latent space and measure the error.

GANs excel at unsupervised anomaly detection because they require only samples of normal, baseline spectrum activity for training. The model learns the manifold of legitimate transmissions—including modulation schemes, bandwidths, and temporal patterns—without needing labeled examples of anomalies. This is critical for detecting unknown or zero-day threats that have never been seen before.

Unlike traditional feature-based methods, GANs can operate directly on raw in-phase and quadrature (I/Q) samples or complex spectrograms. The convolutional layers in a Deep Convolutional GAN (DCGAN) automatically learn hierarchical representations, from low-level phase transitions to high-level modulation patterns. This eliminates the need for manual feature engineering and captures subtle anomalies invisible to hand-crafted statistics like spectral kurtosis.

A standard GAN maps from latent space to data space. A Bidirectional GAN (BiGAN) adds an encoder network that learns the inverse mapping—from data back to latent space. For anomaly detection, this enables reconstruction-based scoring: a normal signal will map to a latent vector that decodes back to a nearly identical signal, while an anomalous signal will have a high reconstruction error. This also helps localize the anomalous time-frequency bins.

GANs can be conditioned on environmental metadata—such as time of day, receiver location, or noise floor estimates—to learn a conditional distribution of normal spectrum behavior. This prevents false alarms caused by diurnal usage patterns or varying propagation conditions. The model learns that a high-power transmission at 3 AM is anomalous, while the same transmission at noon is normal.

The Generator is a neural network trained to synthesize realistic I/Q samples or spectrograms that mimic the statistical distribution of normal, baseline RF activity. Its objective is to produce outputs that the Discriminator cannot distinguish from real, non-anomalous spectrum data. In anomaly detection, a well-trained Generator acts as a model of normality; it learns the manifold of legitimate signals, noise floors, and known interference patterns, but cannot reconstruct or generate novel, anomalous events.

The Discriminator is a binary classifier trained adversarially to distinguish between real spectrum captures and the synthetic data produced by the Generator. During training, it learns the subtle statistical signatures of authentic RF data. In the anomaly detection phase, the Discriminator is repurposed as a feature extractor or direct anomaly scorer. A signal segment that the Discriminator confidently flags as 'fake'—or that produces an anomalous intermediate feature representation—is identified as a deviation from the learned normal distribution.

Adversarial training is the minimax game between the Generator and Discriminator. The Generator minimizes the Discriminator's ability to detect fakes, while the Discriminator maximizes its classification accuracy. This process is formalized as:

• Generator Loss: Maximize the probability of the Discriminator making a mistake.
• Discriminator Loss: Minimize classification error on real vs. fake batches. The equilibrium is reached when the Generator perfectly models the data distribution, and the Discriminator can no longer differentiate, providing a robust model of normality for anomaly detection.