Mahalanobis Distance

Unlike Euclidean distance, Mahalanobis distance is unitless and scale-invariant. It automatically normalizes the data by the variance in each dimension. A difference of 10 units in a feature with high variance contributes less to the distance than a difference of 2 units in a feature with very low variance. This is critical in RF fingerprinting where features like carrier frequency offset and I/Q imbalance operate on vastly different numerical scales.

The metric transforms the feature space into a space where the data distribution is isotropic (spherical) with unit variance. It effectively stretches and rotates the coordinate axes based on the inverse covariance matrix. If two features are highly correlated, the distance metric accounts for this redundancy, preventing the model from double-counting evidence. This is essential for distinguishing tightly clustered authorized emitters from unknown devices.

For a multivariate Gaussian distribution, the squared Mahalanobis distance is directly proportional to the negative log-likelihood of the observation. A small distance corresponds to a high probability density. This property allows open set recognition systems to set a threshold on the Mahalanobis distance to reject outliers; a sample exceeding a critical chi-squared value is statistically flagged as an out-of-distribution or unknown emitter.

The classical Mahalanobis distance is sensitive to outliers because the sample mean and covariance are easily skewed. In adversarial RF environments, a single spoofed signal can distort the boundary. The Minimum Covariance Determinant (MCD) estimator computes the mean and covariance from a subset of 'clean' data, making the distance metric robust against contamination and preventing open space risk from being artificially inflated by poisoned training samples.

In deep metric learning for emitter identification, the Mahalanobis distance is often computed in the feature embedding space rather than the raw input space. A neural network learns a transformation where in-distribution samples follow a compact Gaussian-like distribution. The distance to the class prototype can then be used as a confidence score for open set rejection, often calibrated using Extreme Value Theory (EVT) to model the tail of the distance distribution.

Calculating the Mahalanobis distance requires inverting the covariance matrix, an O(d³) operation for d features. For high-dimensional RF fingerprints, this can be a bottleneck. Practical implementations often use:

• Dimensionality reduction (PCA) to pre-condition the space
• Cholesky decomposition for numerical stability
• Diagonal approximations when features are engineered to be decorrelated This ensures real-time authentication latency for edge-deployed systems.

The straight-line distance between two points in feature space, calculated as the square root of the sum of squared differences. Unlike Mahalanobis distance, Euclidean distance assumes all features are independent and equally scaled, making it blind to correlations. In open set recognition, Euclidean distance fails when features covary—a sample may appear close to a class mean in Euclidean space but be highly anomalous when covariance is considered. It serves as the baseline metric that Mahalanobis improves upon by incorporating the inverse covariance matrix.

A square matrix where each element represents the covariance between two feature dimensions. The diagonal contains variances of individual features, while off-diagonal elements capture linear relationships between feature pairs. In Mahalanobis distance computation, the inverse covariance matrix whitens the feature space, transforming correlated axes into independent, unit-variance dimensions. This ensures that a deviation along a high-variance principal component is penalized less than the same deviation along a low-variance direction.

The probability that a model incorrectly classifies a sample from an unknown class as belonging to a known class. Mahalanobis distance directly addresses open space risk by defining compact, ellipsoidal decision boundaries around each known class. Points far from the training distribution in the Mahalanobis sense are assigned high distance scores and can be rejected. This contrasts with SoftMax-based classifiers, which project unbounded decision regions into open space, creating zones where unknown samples receive high-confidence known-class predictions.

A statistical technique that fits a Weibull distribution to the tail of Mahalanobis distance scores for each known class. This models the probability that a sample at a given distance belongs to that class, enabling calibrated rejection thresholds. The Weibull distribution is particularly suited for extreme value modeling because it captures the behavior of the smallest or largest values in a dataset. In the OpenMax algorithm, Weibull-calibrated Mahalanobis distances replace the standard SoftMax layer to estimate the probability of an unknown class.

Deep Support Vector Data Description trains a neural network to map normal data into a minimal-volume hypersphere in the embedding space. The distance from the sphere center serves as an anomaly score. When combined with Mahalanobis distance, the spherical boundary can be replaced with an ellipsoidal one defined by the covariance structure of the embeddings. This hybrid approach—learning a compact representation with a neural network while using Mahalanobis distance for scoring—provides tighter class boundaries and improved open set rejection compared to either method alone.

A low-dimensional vector representation where semantic similarity is preserved as geometric proximity. Mahalanobis distance is computed in this embedding space rather than on raw input features. Modern open set recognition pipelines use deep networks to learn embeddings where class distributions are approximately Gaussian, making Mahalanobis distance a natural fit. The quality of the embedding directly determines the effectiveness of distance-based rejection—poorly separated or non-Gaussian class clusters degrade Mahalanobis-based detection of unknown emitters.