Counterfactual Explanation

Unlike feature attribution methods that only highlight important variables, counterfactuals provide a direct path to a different outcome. The explanation takes the form: 'If feature X had been value Y instead of Z, the prediction would have changed.'

• Directive: Tells the user exactly what to change
• Contrastive: Answers 'Why P and not Q?'
• Practical: Generates realistic, achievable changes within feasibility constraints

For example, in a loan application denied by an RF signal classifier, a counterfactual might state: 'If the signal-to-noise ratio had been 2.1 dB higher, the emitter would have been classified as authorized.'

A valid counterfactual must be as close as possible to the original instance in feature space. This proximity constraint ensures the explanation is minimally disruptive and realistically achievable. Sparsity further requires that only a small number of features are altered.

• L1 or L2 distance is minimized between the original and counterfactual instance
• Feature count penalty encourages changing only 1-3 variables
• Weighted distances account for differing costs of changing categorical vs. continuous features

This dual constraint prevents trivial or absurd explanations that suggest overhauling every parameter simultaneously.

The generated counterfactual must lie within the data manifold—it must represent a realistic, observable instance, not an adversarial artifact. This is enforced through:

• Density constraints: The counterfactual must reside in a high-density region of the training distribution
• Causal constraints: Changes must respect known causal relationships and immutable characteristics
• Domain-specific guardrails: For RF applications, generated IQ samples must obey physical laws like spectral mask regulations

A counterfactual suggesting an impossible hardware configuration provides no actionable insight and erodes trust in the explanation system.

A single counterfactual may not capture all possible paths to recourse. Diverse counterfactual explanation algorithms generate multiple, meaningfully distinct alternatives, allowing the user to choose the most convenient path.

• Determinantal point processes (DPP) enforce diversity among generated instances
• Clustering-based selection ensures coverage of different regions of the decision boundary
• User preference integration allows filtering by cost, effort, or specific feature constraints

In RF emitter identification, one counterfactual might suggest adjusting transmission power while another suggests shifting carrier frequency—both achieving reclassification but through different mechanisms.

Standard counterfactual search can produce spurious explanations that exploit non-causal correlations. Causal counterfactuals use a structural causal model (SCM) to ensure that suggested changes respect the true data-generating process.

• Intervention-based generation: Changes are modeled as do-operations, not conditional observations
• Immutable feature respect: Protected attributes like hardware serial numbers cannot be altered
• Downstream effect propagation: Changing a root cause variable correctly updates all dependent features

This is critical in RFML where altering one signal parameter causally affects others through known physical relationships.

Counterfactual generation algorithms are themselves vulnerable to manipulation. Adversarial counterfactuals can be crafted to hide biased decision boundaries or suggest unrealistic recourse paths.

• Explanation consistency checks verify that small input perturbations do not yield wildly different counterfactuals
• Robust optimization incorporates worst-case perturbations during counterfactual search
• Auditability metrics quantify the stability of generated explanations over time and across similar instances

For mission-critical RF systems, ensuring that counterfactual explanations remain stable under channel noise and interference is essential for operator trust calibration.