Reconstruction Error

The reconstruction error is typically computed as the Mean Squared Error (MSE) or Mean Absolute Error (MAE) between the input vector x and the reconstructed output x̂.

• MSE: L(x, x̂) = 1/n Σ(xᵢ - x̂ᵢ)² — penalizes large deviations quadratically
• MAE: L(x, x̂) = 1/n Σ|xᵢ - x̂ᵢ| — more robust to outliers in the error distribution
• For complex-valued I/Q data, the error is often computed on the magnitude and phase components separately
• The loss function choice directly impacts which types of anomalies are most detectable

A critical operational parameter is the decision threshold that separates normal signal reconstructions from anomalous ones.

• Static threshold: Set based on the 95th or 99th percentile of reconstruction errors on a validation set of known-normal data
• Dynamic threshold: Adapts to changing noise floors using exponential moving averages or 3-sigma rules on recent error history
• Percentile-based: Flags the top k% of samples in a sliding window as anomalous
• Threshold calibration requires balancing false positive rate (nuisance alerts) against false negative rate (missed intrusions)

The statistical properties of reconstruction errors across a signal batch reveal the nature of the anomaly.

• Uniformly high error across all dimensions suggests a completely novel signal type or modulation scheme
• Sparse high error concentrated in specific frequency bins indicates narrowband interference or a rogue carrier
• Bimodal error distribution may indicate the presence of two overlapping signals where only one was expected
• Temporal bursts of high error followed by recovery suggest pulsed jamming or intermittent unauthorized transmissions

The dimensionality of the latent space directly controls the autoencoder's sensitivity to anomalies.

• A narrow bottleneck forces the model to learn only the most dominant features of normal data, making it highly sensitive to any deviation
• An overly wide bottleneck risks learning the identity function, where even anomalous inputs are reconstructed perfectly
• The optimal latent dimension is typically found through hyperparameter search using reconstruction error on a held-out validation set
• Information-theoretic approaches like the Variational Autoencoder (VAE) impose a prior distribution on the latent space, providing a more principled anomaly score via the evidence lower bound (ELBO)

Advanced architectures compute reconstruction error at multiple temporal and spectral resolutions to catch anomalies at different scales.

• Frame-level error: Computed on individual spectrogram frames for instantaneous anomaly detection
• Sequence-level error: Aggregated across a sliding window of frames to detect slowly developing anomalies like concept drift
• Multi-resolution autoencoders reconstruct the input at multiple scales, with anomalies flagged when errors are inconsistent across scales
• Attention-weighted error uses learned attention maps to focus the anomaly score on the most salient signal components

To move beyond a single scalar anomaly score, gradient attribution methods identify which input features contributed most to the reconstruction error.

• Saliency maps: Compute the gradient of the reconstruction loss with respect to the input, highlighting the frequency bins or time steps most responsible for the error
• Integrated Gradients: A more robust attribution method that integrates gradients along a path from a baseline (e.g., zero signal) to the actual input
• This provides explainability to spectrum operators, showing exactly where in the time-frequency plane the anomaly resides
• Attribution maps can be used to trigger targeted mitigation, such as notching a specific interferer without disrupting the entire band

The foundational architecture that produces reconstruction error. An autoencoder is trained exclusively on normal spectrum data to learn an identity function via a bottleneck latent space. During inference, normal signals are reconstructed with low error, while anomalous or novel signals—unseen during training—produce a high reconstruction error that serves as the anomaly score. Variants include denoising autoencoders and sparse autoencoders.

A temporal extension of the standard autoencoder designed for sequential spectrum data. Long Short-Term Memory layers in both the encoder and decoder capture long-range temporal dependencies in normal signal behavior. The model is trained to reconstruct sequences of normal I/Q samples or spectrograms. A high reconstruction error on a new sequence indicates a temporal anomaly, such as an unexpected transmission pattern or a burst of interference.

A probabilistic alternative to the standard autoencoder. Instead of learning a deterministic latent vector, a VAE learns the parameters of a probability distribution (typically Gaussian) over the latent space. Anomaly detection uses the reconstruction probability rather than raw reconstruction error—a sample is anomalous if its latent representation falls in a low-probability region of the learned distribution. This provides a more principled, probabilistic anomaly score.

The broader task that reconstruction error addresses. OOD detection identifies inputs that differ fundamentally from the training data distribution. In spectrum monitoring, this means detecting novel signal types, unknown modulation schemes, or rogue emitters. Reconstruction error is one OOD scoring function; others include softmax confidence, energy-based scores, and Mahalanobis distance in feature space. OOD detection is critical for open-world spectrum environments where new threats constantly emerge.

A multivariate distance metric that measures how many standard deviations a point is from the mean of a distribution, accounting for covariance between features. Unlike Euclidean distance, it respects the shape of the data distribution. In spectrum anomaly detection, it is often applied in the feature embedding space of a trained model. A high Mahalanobis distance indicates a sample is statistically distant from the normal class, serving as an alternative or complementary anomaly score to reconstruction error.

The identification of changes in the underlying statistical properties of spectrum data over time. While reconstruction error detects point anomalies, concept drift detection identifies gradual or sudden shifts in the normal baseline—such as a new permanent emitter or a change in ambient noise floor. Drift detection algorithms like ADWIN or Page-Hinkley monitor the reconstruction error distribution itself; a sustained increase in mean error signals that the autoencoder's learned normality is becoming stale and requires retraining.