Self-Attention Spectrum Sensing

Unlike static threshold detectors, self-attention computes a learned relevance score for every time-frequency bin in a spectrogram. The model dynamically assigns higher weights to bins containing signal energy while suppressing noise-only regions. This is achieved by computing query, key, and value vectors for each bin, where the attention weight between two bins is proportional to the dot product of their query and key vectors. The result is a context-aware mask that adapts to the instantaneous spectral environment, enabling the detection of signals buried below the noise floor that a simple energy detector would miss.

Traditional convolutional neural networks are limited by their receptive field size—a kernel can only see a fixed local neighborhood of time-frequency bins. Self-attention removes this constraint by allowing every bin to attend directly to every other bin in the spectrogram. This is critical for detecting signals with non-contiguous spectral occupancy, such as frequency-hopping spread spectrum transmissions, or signals exhibiting long temporal correlations, such as radar pulses with low pulse repetition intervals. The attention matrix explicitly models relationships between distant spectral events.

A single attention head might learn to focus on narrowband continuous tones, while another head simultaneously tracks broadband transient pulses. Multi-head self-attention projects the time-frequency tokens into multiple distinct representation subspaces, allowing the model to jointly attend to different types of signal structures in parallel. The outputs from all heads are concatenated and linearly projected, creating a rich, multi-faceted representation of the spectral environment that captures diverse modulation signatures and interference patterns simultaneously.

The self-attention mechanism is permutation-invariant—it has no inherent understanding of the ordering of frequency bins or time steps. To preserve the spectrogram's structure, positional encodings are added to the input tokens. For spectrum sensing, specialized encodings such as Rotary Position Embedding (RoPE) or learned frequency-domain positional encodings inject knowledge of a bin's absolute frequency and relative temporal position. This allows the model to distinguish a signal at 2.4 GHz from an identical signal at 5 GHz, and to understand the sequential nature of time.

After the attention mechanism computes pairwise relevance scores, the output for each time-frequency bin is a weighted sum of all other bins' value vectors, where the weights are the attention scores. This means the representation of a single bin is enriched with information from the entire spectrogram. A weak, intermittent signal component can be reinforced by attending to other occurrences of the same signal across time and frequency, effectively performing non-local signal integration that dramatically improves detection sensitivity in low signal-to-noise ratio (SNR) regimes.

Self-attention spectrum sensing models are often pre-trained using Masked Spectrum Modeling (MSM). Large portions of a spectrogram are randomly masked, and the model is trained to reconstruct the missing time-frequency bins from the visible context. This self-supervised objective forces the model to learn the statistical structure of legitimate signals and background noise without requiring labeled data. The pre-trained encoder can then be fine-tuned on a small labeled dataset for specific detection tasks, achieving high accuracy even with limited annotated examples of rare or novel signal types.