Feature Embedding

Compresses raw I/Q samples or high-dimensional spectral features into a low-dimensional latent space (typically 2-256 dimensions). This removes redundancy and noise while preserving the essential structure needed for downstream tasks.

• Reduces computational complexity for real-time scoring
• Mitigates the curse of dimensionality in distance-based anomaly detectors
• Enables visualization of signal clusters in 2D or 3D space

Maps signals with similar physical characteristics to nearby points in the embedding space. A Wi-Fi beacon and a Bluetooth advertisement—both bursty, short-duration signals—will cluster closer than a continuous-wave tone.

• Euclidean distance or cosine similarity directly reflects signal similarity
• Enables k-Nearest Neighbor retrieval of historically similar events
• Learned similarity often outperforms hand-crafted distance metrics

Learned embeddings are robust to irrelevant variations like carrier frequency offset, minor timing jitter, or thermal noise fluctuations. The network learns to ignore these factors during training.

• A signal shifted by 100 Hz maps to nearly the same vector
• Reduces false positives from benign environmental drift
• Contrast with brittle, hand-tuned feature extractors

Embeddings trained on one task (e.g., modulation recognition) can be repurposed for anomaly detection without retraining the entire network. The latent space captures fundamental signal physics.

• Train once on labeled modulation data
• Reuse the frozen encoder for unsupervised anomaly scoring
• Dramatically reduces the need for labeled anomaly examples

Anomalies are identified by their distance from the centroid of normal embeddings or by their position outside a learned boundary. A rogue emitter with an unseen modulation scheme will map to an isolated region of the latent space.

• Mahalanobis distance accounts for the covariance of normal clusters
• Deep SVDD learns a minimal hypersphere around normal embeddings
• Distance thresholds can be tuned for desired false-alarm rates

For sequential models like LSTM Autoencoders, embeddings capture the temporal dynamics of normal spectrum behavior. A sudden jump in the embedding trajectory signals an anomalous event.

• Embedding vectors evolve smoothly during normal operation
• Abrupt transitions indicate state changes or interference onset
• Enables online anomaly detection on streaming I/Q data

A neural network trained to reconstruct normal signal data through a bottleneck layer, where the reconstruction error serves as the anomaly score. The feature embedding is the compressed representation at the bottleneck. Key characteristics:

• Encoder compresses input into low-dimensional embedding
• Decoder attempts reconstruction from embedding only
• High reconstruction error indicates deviation from learned normality
• Embedding captures essential structure of normal signals

A generative model that learns a probabilistic latent space of normal RF signals. Unlike standard autoencoders, the VAE encodes inputs as mean and variance vectors parameterizing a Gaussian distribution. Anomaly detection uses:

• Reconstruction probability instead of raw error
• Latent space sampling for likelihood estimation
• KL divergence regularization for smooth embeddings
• More robust to noisy inputs than deterministic embeddings

A neural one-class classification method that learns to map normal data into a minimal hypersphere in feature space. The embedding is trained to cluster tightly around a center point:

• Anomalies fall outside the learned hypersphere boundary
• Distance from center serves as anomaly score
• No reconstruction required, only forward encoding
• Effective when normal data is highly structured

A training paradigm where embeddings are learned from unlabeled spectrum data by solving pretext tasks. Common approaches include:

• Contrastive learning: pull similar signal segments together, push dissimilar apart
• Masked prediction: reconstruct intentionally corrupted portions of spectrograms
• Temporal ordering: predict correct sequence of shuffled signal frames
• Embeddings capture semantic structure without manual labeling

The process of applying anomaly detection directly to raw in-phase and quadrature samples, bypassing traditional feature engineering. Embeddings are learned end-to-end from complex-valued time-series data:

• Preserves phase relationships lost in power spectral density
• Convolutional layers extract hierarchical temporal features
• Embedding captures modulation-specific signatures
• Enables detection of subtle waveform anomalies invisible to spectral analysis

The task of identifying inputs that differ fundamentally from the training data distribution. Feature embeddings enable OOD detection through:

• Mahalanobis distance in embedding space
• Density estimation on learned representations
• Energy-based scoring from embedding magnitudes
• Critical for detecting novel signal types in open-world spectrum environments where unknown emitters constantly appear