Deep SVDD

The core principle of Deep SVDD is to train a neural network to map all normal training samples into a feature space such that they are enclosed by the smallest possible hypersphere. The network is trained jointly with the center c of this hypersphere. The objective function penalizes the distance of each mapped point from c, effectively minimizing the volume of the hypersphere. This creates a tight, compact boundary around the learned representation of normality.

Anomaly detection is performed by calculating the distance of a new data point's feature representation from the learned hypersphere center c. The anomaly score is defined as s(x) = ||φ(x; W) - c||², where φ is the neural network with weights W. A point is classified as an anomaly if its distance exceeds the radius R of the hypersphere. This provides a continuous, interpretable score rather than a binary label.

A critical failure mode in Deep SVDD is hypersphere collapse, where the network learns to map all inputs to a constant point, resulting in a trivial solution with zero volume. This is prevented by:

• No bias terms: Removing bias parameters from the network architecture.
• No bounded activations: Avoiding saturating functions like sigmoid or tanh.
• Center c fixed: Computing the center as the mean of an initial forward pass and keeping it fixed, rather than learning it as a parameter.

The Soft-Boundary Deep SVDD variant allows some training points to fall outside the hypersphere, controlled by a hyperparameter ν (nu). This trades off between the volume of the hypersphere and the number of permitted outliers. The objective becomes: minimize R² + (1/νn) Σ max(0, ||φ(x_i) - c||² - R²). This is essential when the training data contains unknown contamination or noise.

Unlike autoencoder-based anomaly detection, which relies on reconstruction error, Deep SVDD directly optimizes a compact boundary in feature space. This avoids the assumption that anomalies will always reconstruct poorly. Deep SVDD is often more sample-efficient for one-class problems because it does not waste capacity learning to reconstruct the input, focusing entirely on the discriminative boundary around normal data.

In spectrum anomaly detection, Deep SVDD is applied to raw I/Q samples or spectrogram features to identify rogue emitters. The model is trained exclusively on ambient background noise and known licensed transmissions. Any new signal—such as an unauthorized transmitter or jamming waveform—maps to a point far from the learned center, triggering an alert. This is critical for LPI detection and open set recognition in electronic warfare.

A classical machine learning algorithm that defines a decision boundary around normal signal features in a high-dimensional space. Unlike Deep SVDD's learned hypersphere, One-Class SVM uses a kernel trick to map data and find a hyperplane that separates normal points from the origin. It is highly effective for small, clean datasets but struggles with the high-dimensional raw I/Q data that neural methods handle natively.

An ensemble method that explicitly isolates anomalies by randomly partitioning data. It exploits the property that anomalous points are rare and different, thus requiring fewer splits to be separated from the rest of the dataset. Unlike Deep SVDD, it does not learn a compact representation of normality but is exceptionally fast and effective for high-dimensional tabular features extracted from spectrum scans.

A neural network trained to reconstruct normal signal data. Anomalies are identified by a high reconstruction error on new inputs. While Deep SVDD directly optimizes a compact hypersphere in feature space, autoencoders learn a latent manifold of normality. Variants like LSTM Autoencoders are particularly suited for temporal spectrum data, capturing sequential patterns that a static hypersphere might miss.

The overarching task of identifying inputs that differ fundamentally from the training data distribution. Deep SVDD is a specific OOD detection technique. Other methods include Maximum Softmax Probability for classifiers and Mahalanobis distance in feature space. In open-world spectrum environments, OOD detection is critical for identifying novel signal types and unauthorized emitters without prior examples.

A density-based algorithm that identifies anomalous data points by measuring the local deviation of a given sample's density relative to its neighbors. Unlike Deep SVDD's global boundary, LOF can detect local anomalies that might be hidden within a global hypersphere. It is particularly useful for identifying subtle interference patterns that are only abnormal relative to their immediate spectral neighborhood.

A learned, low-dimensional vector representation of raw I/Q samples or spectral features. Deep SVDD relies on a neural network to learn these embeddings such that normal data clusters tightly. The quality of the embedding is paramount. Techniques like contrastive learning or self-supervised pretraining can significantly improve the compactness of the normal class representation before applying the SVDD objective.