Explainable AI (XAI) for Interference

A game-theoretic approach that assigns each input feature—such as a specific frequency bin or time segment—an importance value for a particular classification. SHAP values quantify how much each element of an IQ sample or spectrogram contributed to pushing the model toward a 'barrage jammer' versus 'protocol-aware jammer' decision.

• Based on Shapley values from cooperative game theory
• Provides both global model interpretability and local per-prediction explanations
• Computationally intensive for high-dimensional RF data; optimized variants like Kernel SHAP or Deep SHAP are used
• Example: Reveals that a classifier relies on a narrow 2 MHz spur rather than broadband noise to identify a specific jammer type

A visualization technique that highlights which pixels in a time-frequency representation most influenced a neural network's classification decision. Saliency maps compute the gradient of the output class score with respect to the input spectrogram, producing a heatmap overlay.

• Identifies temporal-spectral regions critical for classification
• Variants include Grad-CAM and SmoothGrad for less noisy visualizations
• Enables analysts to verify that models focus on physically meaningful signal structures rather than artifacts
• Example: A saliency map confirms that a model distinguishing QPSK from QAM-16 is attending to phase transition regions, not background noise

A perturbation-based method that approximates a complex RF classifier locally with an interpretable surrogate model. LIME generates synthetic samples by masking portions of the input signal and observing output changes, then fits a simple linear model to explain the decision boundary around a specific prediction.

• Model-agnostic: works with any black-box classifier
• Perturbs superpixels in spectrograms or contiguous time-frequency blocks
• Provides counterfactual insights: which signal features, if removed, would change the classification
• Example: Demonstrates that removing a periodic 10 ms pulse eliminates a 'reactive jammer' classification, confirming temporal pattern dependence

A path-based attribution method that satisfies the completeness axiom, ensuring that feature importance scores sum to the difference between the model's output and a neutral baseline. For RF applications, Integrated Gradients accumulates gradients along a straight-line path from a baseline (typically zero signal or noise) to the actual IQ input.

• Particularly suited for complex-valued neural networks processing raw IQ samples
• Avoids saturation problems common with simple gradient-based saliency
• Identifies which I and Q components at specific time indices drive classification
• Example: Reveals that the quadrature component at sample offset 2048 is the primary discriminator between two modulation schemes

A technique specific to transformer-based signal classifiers that renders the self-attention weight matrices as interpretable heatmaps. Each attention head learns to focus on different temporal or frequency relationships within the input sequence.

• Reveals long-range dependencies the model exploits for interference recognition
• Multi-head attention can be inspected to see which heads specialize in periodic patterns, frequency hops, or transient events
• Helps validate that the model learns physically plausible signal structure
• Example: Attention maps show a model tracking a frequency-hopping jammer by attending to correlated energy shifts across distant time steps

A method that tests whether a neural network has learned human-understandable concepts relevant to RF classification. CAVs are derived by training a linear classifier to distinguish between examples containing a concept (e.g., 'frequency-modulated component') and random counterexamples, then measuring how sensitive the model's predictions are to that concept direction in activation space.

• Enables concept-based explanations rather than raw feature attribution
• Tests for concepts like 'chirp presence,' 'pulse periodicity,' or 'phase discontinuity'
• Quantifies with TCAV scores how important a concept is for a class
• Example: Confirms that a 'linear FM jammer' classifier has internalized the concept of a linearly sweeping frequency tone

The core mathematical techniques that power XAI, quantifying the contribution of each input feature to a model's prediction.

• SHAP (SHapley Additive exPlanations): Uses cooperative game theory to assign fair credit to each time-frequency bin or statistical feature.
• LIME (Local Interpretable Model-agnostic Explanations): Approximates the complex RF classifier locally with a simpler, interpretable model.
• Saliency Maps: Highlight the specific pixels in a spectrogram or regions in an IQ sample that most activated the neural network's decision.
• Integrated Gradients: Computes the path integral of gradients from a baseline (e.g., pure noise) to the actual input, satisfying the completeness axiom.

Standard neural networks process real numbers, but RF signals are inherently complex (I/Q data). CVNNs preserve this structure by using complex-valued weights, activations, and backpropagation. This is critical for XAI because interpreting a model that has already destroyed phase information is misleading. XAI methods for CVNNs must explain contributions to both magnitude and phase, providing a physically meaningful attribution. For example, an explanation might reveal that the classifier is keying on a specific phase rotation pattern unique to a particular jammer's power amplifier non-linearity.

Traditional classifiers assume all test signals belong to one of the classes seen during training. Open-Set Recognition builds models that can say 'I don't know' when encountering a novel, never-before-seen interference waveform. XAI is vital here to justify why a signal was rejected. Instead of a classification label, the explanation might show that the signal's cyclostationary features or higher-order statistics fall far outside the known distribution, giving the spectrum analyst actionable intelligence about a potential new threat rather than a silent misclassification.