Grad-CAM

A visualization technique that modifies the standard backward pass through ReLU activation functions. While Grad-CAM provides coarse localization, Guided Backpropagation sharpens the result by only backpropagating positive gradients through positive activations, preventing negative contributions from flowing backward. This produces fine-grained, high-resolution saliency maps. The combination of Grad-CAM and Guided Backpropagation creates Guided Grad-CAM, which yields both high-resolution and class-discriminative visualizations. In RF applications, this pinpoints the exact temporal samples within an IQ sequence that triggered a modulation classification.

A foundational interpretability tool that visualizes the influence of each input feature on a model's prediction by computing the gradient of the target class score with respect to the input. Unlike Grad-CAM, which is restricted to convolutional layers, vanilla saliency maps work on any differentiable architecture. Key characteristics:

• Noisy and high-frequency: Raw saliency maps often appear grainy
• Input-space attribution: Highlights individual pixels or IQ samples
• Model-agnostic: Applicable to any differentiable model For RF signals, a saliency map might highlight specific transient regions in a burst transmission that the network uses for emitter identification.

The direct predecessor to Grad-CAM that produces localization maps by projecting the weights of the final global average pooling layer back onto the last convolutional feature maps. CAM requires a specific architecture where the final layer is global average pooling followed by a single fully-connected layer. This structural constraint limits its applicability. Grad-CAM generalizes CAM to any CNN architecture by using gradient information instead of learned weights. In wireless signal processing, CAM was initially applied to spectrogram classification to identify which frequency components drove a signal type prediction.

A perturbation-based enhancement of Grad-CAM that eliminates the dependency on gradients entirely. Instead of using gradients as weights, Score-CAM measures the increase in confidence for the target class when each activation map is used as a spatial mask on the input. This approach:

• Avoids gradient saturation issues common in deep networks
• Provides more faithful attributions by directly measuring causal impact
• Is computationally heavier due to multiple forward passes For RF anomaly detection, Score-CAM can more reliably identify the specific spectral anomaly that triggered an alert without being misled by vanishing gradients in saturated non-linear regions.

A decomposition-based explanation method that redistributes a neural network's prediction score backward through the network using a conservation property. Unlike gradient-based methods like Grad-CAM, LRP guarantees that the total relevance assigned to all input features exactly equals the output prediction score. This makes LRP particularly suitable for:

• Regulatory compliance where exact accountability is required
• Deep architectures where gradients may shatter
• Quantitative analysis of feature contributions In RF fingerprinting, LRP can decompose a classification decision to show exactly which hardware impairment signatures contributed what proportion to the final emitter identification.

An axiomatic attribution method that satisfies Sensitivity and Implementation Invariance—two fundamental properties that many gradient-based methods violate. It computes feature importance by accumulating gradients along a straight-line path from a baseline input (e.g., a zero signal or noise floor) to the actual input. Key advantages:

• Path integral captures non-linear effects missed by single-point gradients
• Baseline selection allows domain-specific prior knowledge injection
• Satisfies completeness: attributions sum to the output difference For spectrum sensing, the baseline can be set to a clean noise floor, revealing which signal components the model uses to distinguish a radar pulse from background interference.