SEI Model Explainability

Saliency maps compute the gradient of the emitter classification score with respect to each input sample in the I/Q sequence. This highlights which specific time-domain samples most influence the model's decision.

• Mechanism: Backpropagation of the target class score to the input layer.
• RF Application: Identifies if the model focuses on the turn-on transient, preamble, or specific steady-state modulation artifacts.
• Limitation: Standard saliency maps can be noisy; SmoothGrad is often applied by averaging gradients over multiple noise-injected copies of the signal to produce visually coherent heatmaps.

Gradient-weighted Class Activation Mapping (Grad-CAM) is adapted for RF by applying it to the final convolutional layers of a model that processes time-frequency spectrograms.

• Output: A coarse, class-discriminative heatmap overlaid on the spectrogram, revealing which frequency bands and time intervals are critical for identification.
• Use Case: Distinguishing whether the model relies on the carrier frequency offset artifact, phase noise sidebands, or power amplifier non-linearity patterns.
• Architecture: Requires a CNN backbone where the final convolutional feature maps retain spatial correspondence to the input spectrogram.

Guided Backpropagation modifies the standard backpropagation of ReLU activations to only propagate positive gradients, suppressing negative contributions.

• Result: Produces high-resolution, fine-grained visualizations of the exact signal features that activate specific neurons.
• RF Benefit: When combined with Grad-CAM (Guided Grad-CAM), it provides both the high-resolution detail of which I/Q samples matter and the class-discriminative localization of which time-frequency regions are important.
• Interpretation: Sharpens the visualization of I/Q imbalance artifacts or clock jitter signatures that are invisible in standard spectrograms.

LRP decomposes the model's output prediction backwards through the network using a conservation principle, assigning a relevance score to each input feature.

• Advantage: Unlike gradient-based methods, LRP does not suffer from shattered gradients in very deep networks and provides theoretically grounded attribution.
• RF SEI Application: Identifies whether the model's decision is based on genuine hardware impairments or spurious channel artifacts, which is critical for validating channel-robust fingerprinting claims.
• Variants: LRP-αβ rules allow tuning the balance between positive and negative evidence propagation.

SHapley Additive exPlanations (SHAP) applies cooperative game theory to assign each input feature an importance value for a specific prediction.

• Method: Computes the marginal contribution of each time-frequency bin or handcrafted feature (e.g., EVM, I/Q offset) averaged over all possible feature subsets.
• RF Context: Quantifies the relative importance of phase noise versus amplifier non-linearity in distinguishing two specific emitters.
• Computational Cost: Exact SHAP is exponential; Kernel SHAP or Deep SHAP provide efficient approximations for deep RF models.

Testing with CAVs (TCAV) measures the sensitivity of an emitter classifier to high-level, human-interpretable concepts rather than raw input features.

• Process: Define a concept (e.g., "high phase noise" or "severe I/Q imbalance") using a set of example signals. A linear classifier is trained to separate the concept examples from random signals in the model's activation space.
• Output: The TCAV score quantifies how important that abstract concept is to the model's classification of a specific emitter.
• Value: Bridges the gap between raw signal attribution and the physical hardware impairments that domain experts understand.