Mahalanobis Distance

The core advantage of Mahalanobis distance is its use of the covariance matrix of the dataset. This scales the distance calculation by the variance of each feature and the correlation between features. In RF anomaly detection, this means a deviation in a normally stable feature (low variance) is weighted more heavily than the same absolute deviation in a naturally noisy feature (high variance). This prevents high-variance background noise from masking critical anomalies in stable signal parameters.

The squared Mahalanobis distance follows a chi-squared distribution when the underlying data is multivariate normal. This provides a direct statistical method for anomaly thresholding:

• A distance value can be converted to a p-value.
• An observation with a p-value below a set significance level (e.g., < 0.001) is flagged as an outlier. This probabilistic grounding allows operators to set thresholds based on a desired false alarm rate rather than arbitrary distance cutoffs.

Euclidean distance treats all feature axes as orthogonal and equally scaled, creating a circular decision boundary. Mahalanobis distance creates an ellipsoidal decision boundary that molds to the data's natural shape. For example, if a signal's center frequency and bandwidth are highly correlated in normal operation, a point with a typical frequency but atypical bandwidth for that frequency will be correctly identified as anomalous, whereas a Euclidean metric might miss it.

Classical Mahalanobis distance is sensitive to outliers in the training data, which can distort the mean and covariance estimates. In spectrum environments with unknown intermittent interference, a Minimum Covariance Determinant (MCD) or other robust estimator must be used. These methods compute the mean and covariance on a subset of the data that minimizes the determinant of the covariance matrix, ensuring the baseline normality model is not corrupted by the very anomalies it seeks to detect.

In a cognitive radio pipeline, the Mahalanobis distance is applied to a vector of extracted features per signal burst:

• Feature Vector Example: [Spectral Kurtosis, Center Frequency Offset, Symbol Rate Deviation, Phase Noise Variance]
• The distance is computed against a pre-calculated, robust mean and inverse covariance matrix.
• Because the inverse covariance is pre-computed, the online calculation is a fast matrix-vector multiplication, suitable for real-time FPGA or SDR processing.

When the number of feature dimensions approaches or exceeds the number of samples, the covariance matrix becomes singular and non-invertible. This is common in high-dimensional RF fingerprinting. Solutions include:

• Shrinkage estimators that regularize the covariance matrix toward a diagonal target.
• Dimensionality reduction via PCA to project features into a lower-dimensional space before applying the distance metric.
• Using the pseudo-inverse of the covariance matrix as a fallback.

Unlike Euclidean distance, which treats all signal features independently, the Mahalanobis distance accounts for inter-feature correlations. It transforms the feature space by the inverse of the covariance matrix, effectively normalizing the data. A signal's anomaly score is the distance from the centroid of the normal class in this transformed space.

• Formula: D_M(x) = sqrt((x - μ)^T Σ^{-1} (x - μ))
• Thresholding: A high distance indicates a low probability of belonging to the normal distribution.
• Unitless: The result is a dimensionless scalar, simplifying threshold setting across diverse feature sets.

In spectrum monitoring, the Mahalanobis distance is used to distinguish between natural noise floor variations and structured interference. The normal ambient environment is modeled as a multivariate Gaussian distribution of spectral features (e.g., power levels across sub-bands, spectral flatness, kurtosis).

• Baseline Model: The mean vector (μ) and covariance matrix (Σ) define the quiet environment.
• Detection: A sudden increase in the Mahalanobis distance signals the onset of interference, even if the raw power increase is subtle.
• Classification: Different types of interference (e.g., CW tone vs. modulated jammer) will project to different high-distance regions, aiding in categorization.

The effectiveness of the Mahalanobis distance depends critically on the feature space. Raw I/Q samples are often too high-dimensional and noisy. Instead, a compact set of handcrafted or learned features is used.

• Handcrafted Features: Spectral kurtosis, cyclostationary features, and higher-order statistics (HOS) provide a low-dimensional, statistically meaningful input.
• Learned Features: A deep autoencoder can compress raw spectrum data into a compact latent vector. The Mahalanobis distance is then computed in this learned feature space, combining deep learning's representational power with classical statistical rigor.
• Covariance Estimation: A robust estimate of Σ (e.g., using Minimum Covariance Determinant) is essential to prevent outliers from corrupting the model of normality.

The RF environment is non-stationary; the statistical definition of 'normal' changes over time. The Mahalanobis distance is a powerful tool for detecting this concept drift. By continuously computing the distance of a sliding window of recent observations from the established baseline, a system can monitor for gradual or sudden environmental shifts.

• Gradual Drift: A slow, monotonic increase in the average distance indicates the baseline model is becoming stale.
• Sudden Shift: An abrupt, sustained jump in distance signals a major environmental change, such as a new permanent transmitter coming online.
• Model Retraining Trigger: This distance metric can serve as the automated trigger to retrain or adapt the anomaly detection model, ensuring continued accuracy.

For real-time spectrum awareness, the computation must be highly optimized. The core operation is a matrix-vector multiplication. The inverse covariance matrix (Σ^{-1}) is precomputed during the training phase, making the online scoring operation extremely fast.

• Precomputation: The heavy linear algebra (matrix inversion) is done offline.
• Online Scoring: Each new signal vector requires only a subtraction and two matrix multiplications.
• Hardware Suitability: This low computational load makes it ideal for deployment on FPGAs and edge AI processors for instantaneous, on-sensor anomaly flagging.

A learned, low-dimensional vector representation of raw I/Q samples or spectral features. Mahalanobis Distance is often calculated in this embedding space rather than on raw data, as the embedding captures the essential, non-linear characteristics of normal signals. By computing the distance of a new embedding from the centroid of normal embeddings, accounting for the covariance of the cluster, you achieve a robust anomaly score that is sensitive to semantic deviations invisible in the original high-dimensional input space.

A support vector machine algorithm that defines a tight, soft boundary around normal signal features in a high-dimensional space. While a One-Class SVM typically uses a kernelized distance, the Mahalanobis Distance provides a powerful alternative for pre-processing or kernel construction. By using a Mahalanobis kernel, the SVM's boundary becomes sensitive to the covariance structure of the data, creating an ellipsoidal decision surface that more accurately encapsulates the normal class and rejects points outside this multivariate distribution.

The task of identifying inputs that differ fundamentally from the training data distribution. Mahalanobis Distance is a state-of-the-art baseline for OOD detection in neural networks. Instead of using the softmax probability, a class-conditional Gaussian distribution is fitted to the feature vectors of the penultimate layer for each class. An input's minimum Mahalanobis distance to any class centroid serves as the OOD score, providing a more reliable and calibrated measure of novelty than maximum softmax probability.

A density-based algorithm that identifies anomalies by measuring the local deviation of a sample's density relative to its neighbors. Standard LOF uses Euclidean distance, but substituting Mahalanobis Distance creates a variant robust to local feature correlation. This prevents the algorithm from being misled by a dense, highly correlated cluster of normal points that might appear sparse in Euclidean space. The Mahalanobis LOF correctly identifies the true local density structure, flagging points that are outliers within their correlated neighborhood.

A neural one-class classification method that learns to map normal data into a minimal hypersphere in feature space. The standard Deep SVDD uses the Euclidean distance to the sphere's center as the anomaly score. A powerful extension replaces this with the Mahalanobis Distance to the center, learned jointly with the network. This 'Deep SAD' (Deep Semi-supervised Anomaly Detection) approach allows the model to learn a feature space where the normal class is not just compact, but also respects the inherent covariance structure, leading to tighter, more accurate boundaries.