Energy-Based Models (EBM)

The fundamental mechanism of an EBM is learning a function E(x): R^D → R that maps high-dimensional input data (like IQ samples) to a single, unnormalized scalar energy value. Training shapes the energy landscape so that in-distribution data occupies low-energy basins, while out-of-distribution samples are pushed to high-energy regions. Unlike probabilistic models, EBMs do not require explicit normalization via a partition function, making them computationally tractable for complex, high-dimensional signal data.

EBMs are often trained using Contrastive Divergence (CD) or its persistent variant (PCD) to approximate the gradient of the log-likelihood without computing the intractable partition function. The process involves:

• Positive Phase: Lowering energy on real, observed data samples from the training distribution.
• Negative Phase: Raising energy on generated or sampled 'fantasy' particles that represent the model's current belief about the data space. This push-pull dynamic carves the energy landscape to separate known emitter signatures from the surrounding open space.

EBMs natively support open set recognition through a simple thresholding mechanism on the energy score. During inference, an input signal is passed through the energy function. If E(x) > τ (a calibrated threshold), the sample is rejected as unknown or anomalous. This eliminates the need for a separate out-of-distribution detector. The threshold τ is typically calibrated on a held-out validation set to achieve a desired true positive rate, often using Extreme Value Theory (EVT) to model the tail distribution of in-distribution energies.

To generate negative samples during training, EBMs commonly employ Stochastic Gradient Langevin Dynamics (SGLD). This MCMC method iteratively refines a random noise sample by following the gradient of the energy function: x_{t+1} = x_t - (λ/2) ∇E(x_t) + ε, where ε is injected Gaussian noise. The noise prevents the sampler from collapsing to a single mode. For RF fingerprinting, this allows the model to explore the boundary of known device signatures, strengthening the energy barrier against unknown emitters.

A powerful property of EBMs is that multiple energy functions can be additively combined to represent complex, factorized constraints. For emitter recognition, separate EBMs can model distinct signal characteristics—such as cyclostationary features and IQ constellation distortion—and their energies summed: E_total(x) = E_cyclo(x) + E_IQ(x). This compositional nature allows for modular, interpretable models that combine evidence from heterogeneous signal processing pipelines without retraining a monolithic architecture.

EBMs provide a natural measure of epistemic uncertainty through the local geometry of the energy landscape. Inputs that lie in flat, high-energy plateaus indicate high uncertainty and likely unknown classes, while samples in steep, low-energy wells indicate confident, known classifications. Techniques like Monte Carlo Langevin sampling around a test point can estimate the local energy variance, offering a more nuanced rejection criterion than a single threshold. This is critical for distinguishing between noisy known emitters and truly novel devices.