Self-Supervised Learning

The core mechanism of SSL is the pretext task, a carefully engineered objective that forces the model to learn meaningful signal structure without human annotation.

• Contrastive Learning: The model learns to pull representations of augmented versions of the same signal segment together while pushing apart representations of different segments. This teaches the model to identify what makes a signal unique.
• Masked Signal Modeling: Random portions of a spectrogram or I/Q sequence are masked, and the model must reconstruct the missing content from context, learning the statistical regularities of normal transmissions.
• Temporal Order Verification: The model is trained to determine if a sequence of spectrum snapshots is in the correct chronological order, forcing it to learn the lawful temporal dynamics of the RF environment.

Once the pretext task is solved, the model has produced a feature embedding—a compact, high-dimensional vector that captures the essential characteristics of normal spectrum behavior.

• Downstream Application: These learned embeddings become the input to a lightweight anomaly detector, such as a One-Class SVM or a simple Mahalanobis Distance calculation, which defines the boundary of normality.
• Open Set Recognition: Because the SSL model was not trained on a closed set of emitter classes, its representations generalize to identify any deviation from learned normality, making it ideal for detecting rogue emitters and LPI signals.
• Feature Transferability: The representations learned via SSL on one frequency band often transfer effectively to adjacent bands, reducing the need to retrain from scratch for each monitored allocation.

A powerful SSL architecture for sequential spectrum data is Contrastive Predictive Coding. It learns by predicting future latent representations from past ones.

• Mechanism: A convolutional encoder compresses raw I/Q samples into a sequence of latent vectors. An autoregressive model, typically a GRU or LSTM, summarizes the past context and predicts future latent vectors. The model is trained to maximize mutual information between the prediction and the actual future observation.
• Spectrum Application: CPC excels at learning the temporal dynamics of spectrum occupancy, including the periodic beaconing of primary users and the bursty nature of packet-switched traffic. An anomaly is detected when the model's prediction of future spectrum state diverges sharply from the observed state.
• Advantage over Autoencoders: Unlike autoencoders that learn to reconstruct every detail, CPC learns a 'slow feature' representation that captures high-level, slowly varying context, making it robust to high-frequency noise.

The electromagnetic environment is non-stationary; the definition of 'normal' changes with time of day, new infrastructure, and seasonal propagation. SSL provides a framework for online anomaly detection that adapts without catastrophic forgetting.

• Continual Pretext Learning: The model continues to solve its pretext task on a rolling window of recent spectrum data, allowing its internal representation of normality to slowly track environmental changes.
• Drift as Anomaly: A sudden, persistent change in the loss of the pretext task itself can serve as a high-level concept drift detection signal, indicating a fundamental shift in the RF environment, such as the activation of a new base station or a persistent jammer.
• Memory Replay: A buffer of representative past signal examples is maintained and interleaved with new data during continual training to prevent the model from overfitting to transient anomalies and forgetting long-term normal patterns.

The engineered self-supervised objective that forces a model to learn meaningful representations from unlabeled data. In spectrum analysis, common pretext tasks include:

• Temporal Order Verification: Determining if a sequence of I/Q samples is in the correct chronological order.
• Signal Rotation Prediction: Predicting the rotation applied to a constellation diagram.
• Context-Aware Encoding: Masking portions of a spectrogram and training the model to reconstruct the missing time-frequency bins. The choice of pretext task directly determines which invariant features the encoder learns, such as modulation shape or symbol rate.

A self-supervised approach that learns representations by pulling positive pairs (augmented views of the same signal) together in embedding space while pushing negative pairs (different signals) apart. Key frameworks adapted for RF data include:

• SimCLR: Maximizes agreement between differently augmented versions of the same spectrogram via a contrastive loss in the latent space.
• MoCo: Maintains a dynamic dictionary of negative samples as a queue, enabling large-batch contrastive training on spectrum datasets.
• BYOL: Bootstraps representations without requiring negative pairs, using a target network to prevent representational collapse when training on I/Q samples.

A reconstruction-based pretext task where a high proportion of input patches (e.g., time-frequency bins in a spectrogram) are randomly masked, and the model learns to reconstruct the missing content. This forces the encoder to learn global semantic understanding of spectrum structure. Key properties:

• High Masking Ratio: Typically 75-80% of spectrogram patches are removed, preventing trivial interpolation.
• Asymmetric Architecture: A heavy encoder processes only visible patches, while a lightweight decoder reconstructs the full signal.
• Anomaly Application: Reconstruction error on unmasked regions indicates deviation from learned normality.

The fundamental goal of self-supervised learning: transforming raw, high-dimensional spectrum data into a compact, structured feature embedding where semantically similar signals cluster together. Effective representations exhibit:

• Modulation Invariance: Signals with the same modulation scheme map to nearby vectors regardless of noise or channel effects.
• Temporal Coherence: Adjacent time slices of a stable transmission maintain consistent embeddings.
• Anomaly Separability: Novel or unauthorized signal types map to low-density regions of the embedding space, enabling detection via Mahalanobis distance or Local Outlier Factor scoring.

An advanced self-supervised paradigm that moves beyond pixel-level reconstruction to predict representations in abstract embedding space. Instead of generating raw spectrogram pixels, JEPA predicts the latent representation of masked signal regions from visible context. Advantages for spectrum data:

• Noise Robustness: Ignores unpredictable high-frequency noise, focusing on semantic signal structure.
• Computational Efficiency: Eliminates the need for a generative decoder operating in high-dimensional input space.
• Energy-Based Models: Often trained with energy functions that assign low energy to compatible signal-context pairs and high energy to incompatible ones.

The process of leveraging frozen self-supervised representations for anomaly detection without fine-tuning. Common scoring methods applied to the learned embeddings include:

• k-Nearest Neighbors (k-NN): Anomaly score is the average distance to the k nearest normal training samples in embedding space.
• Gaussian Mixture Model (GMM): Fits a probability distribution to normal embeddings; low log-likelihood indicates an anomaly.
• One-Class SVM: Learns a tight boundary around normal embeddings; points outside are flagged.
• Deep SVDD: Trains a secondary network to enclose normal embeddings in a minimal hypersphere, with anomalies falling outside the radius.