Perturbation-Based Faithfulness Metric

• Perturbation-based faithfulness metrics are techniques that measure the consistency between explanation-assigned feature importance and actual model output changes under systematic input perturbations.
• They involve diverse methods including feature-level, adversarial, and hidden-space perturbations to assess how individual features or concepts causally influence predictions in vision and language models.
• Key algorithms like AOPC, sufficiency, and c-Eval provide actionable evaluations of explanation quality despite challenges in normalization and out-of-distribution effects.

A perturbation-based faithfulness metric quantifies the degree to which model explanations accurately reflect the input features or concepts that causally influence a model’s predictions, through controlled model responses to systematic, localized input perturbations. This family of metrics forms the empirical backbone of post-hoc interpretability research across feature attribution, concept-based explanations, and explanation faithfulness evaluation in both vision and LLMs.

1. Principles and Taxonomy of Perturbation-Based Faithfulness

Perturbation-based faithfulness metrics evaluate explanations by algorithmically altering input features or internal representations, then observing changes in model outputs or explanations. Faithfulness, in this context, is operationalized as the consistency between the purportedly important features (as given by an explainer) and the actual causal influence of those features as revealed by model behavior under perturbation.

There are three major categories of perturbation-based faithfulness metrics:

• Feature-level input perturbation: Modifications to specific input features (e.g., masking tokens, replacing pixels) assess the output sensitivity attributable to those features. This includes leave-one-out, AOPC, PGI, and c-Eval metrics (Vu et al., 2019, Barr et al., 2023, Gajewski et al., 2024, Edin et al., 2024, Chan et al., 2022).
• Adversarial or semantic perturbation: Input perturbations are crafted to generate minimal or semantic-preserving adversarial examples, with faithfulness defined by tracking explanation changes (e.g., Adversarial Sensitivity) (Manna et al., 2024).
• Hidden-space and concept perturbation: Rather than input features, unit or concept activations in model internals are manipulated to score their causal impact on output decision-making (Li et al., 2024).

A typical metric comprises four elements: (i) the type of perturbation applied, (ii) the operationalization of the model’s target output, (iii) an explanation object (feature subset, attribution ranking, or concept vector), and (iv) an aggregation/statistical procedure to summarize the causal relationship.

2. Classical Algorithms: Input Feature Perturbation

Fundamental metrics such as sufficiency, comprehensiveness, area-over-perturbation-curve (AOPC), decision-flip counts, and prediction gap are realized by sequentially perturbing feature subsets in descending (or ascending) importance order and measuring the resulting shift in prediction (Chan et al., 2022, Edin et al., 2024):

• Comprehensiveness:

$$\operatorname{Comp}(x, S) = p_{c(x)}(x) - p_{c(x)}(x \setminus S)$$

Quantifies the output drop upon removing top-ranked features S; larger values mean greater faithfulness.

• Sufficiency:

$$\operatorname{Suff}(x, S) = p_{c(x)}(x) - p_{c(x)}(x_{S})$$

Measures the extent to which a feature subset alone suffices; lower is better.

• AOPC:

$$\mathrm{AOPC}(f,x,r) = \frac{1}{N}\sum_{i=1}^{N}[f(x) - f(p(x, r_{1:i}))]$$

Provides a cumulative measure across an entire perturbation curve (Edin et al., 2024).

• Prediction Gap on Important Features (PGI/PGI²):

$$\operatorname{PGI}(x, f, a; k) = \frac{1}{m} \sum_{j=1}^m |f(x) - f(\tilde{x}^{(j)})|$$

or with squared loss:

$$\operatorname{PGI}^2(x, \pi, k) = \mathbb{E}_{x'}[(f(x')-f(x))^2]$$

The average output shift when randomly perturbing the k most important features (Barr et al., 2023, Gajewski et al., 2024).

• c-Eval:

$$c_{f,x}(e_x) = \inf \{\| \delta \|_p: \delta_i = 0 \ \forall i \in e_x, \ \arg\max_j f_j(x+\delta) \neq \ell_0 \}$$

Minimal perturbation on non-explanatory features needed to flip the model’s prediction, with explanatory features held fixed (Vu et al., 2019).

All algorithms require specification of perturbation operators (masking, additive noise, etc.), ranking measures (from explainers), and aggregation (averaging over examples or ablated feature counts).

3. Advanced and Composite Faithfulness Metrics

Recent work has introduced more sophisticated compositional and black-box faithfulness frameworks:

• Perturbation-based Faithfulness for LLMs with Self-Explanations:

For input $(q, c, e)$, find minimal context spans $SR$ (sufficient regions) where $M(q,s_j)$ yields the correct answer, and within these, token groups $NK_s$ whose ablation causes error. The final metric,

$$f = \max_{s \in SR} \frac{f_{SR}(s) + f_{NK}(s)}{2}$$

compares overlap between ground-truth sufficient/necessary regions and model-provided self-explanation keywords (Fragkathoulas et al., 2024).

• Concept-based Faithfulness Metrics:

Perturb hidden representation $h$ in the direction of a concept $a$, then quantify $\delta(g(h), g(\xi(h,a)))$, where $\xi$ implements ablation or maximal activation. Aggregated over positions and difference measures (e.g., KL-divergence), this yields $\gamma(a, \xi, \delta)$—how much the concept moves model outputs (Li et al., 2024).

• Adversarial Sensitivity:

Given adversarially-induced model output flip $f(x)\neq f(x')$, the distance $d(W_{x,f}, W_{x',f})$ in explanation space (e.g., 1 minus generalized Kendall’s $\tau$) scores the faithfulness with which the explainer tracks the model’s changed reasoning. High sensitivity is necessary for faithfulness (Manna et al., 2024).

4. Evaluation, Comparison, and Calibration

Empirical evaluation of perturbation-based faithfulness metrics reveals several consensus findings and challenges:

• Diagnosticity and Robustness: Metrics such as comprehensiveness, sufficiency, AOPC, PGI, and ABC all show variable diagnosticity; comprehensiveness and sufficiency are consistently top performers relative to random or adversarially fabricated attributions, with low computational cost (Chan et al., 2022).
• Disagreement across Metrics: No single metric is universally agreed upon or fully robust. ABC is strictly monotonic with respect to ground truth in synthetic linear settings, whereas PGI can be non-monotonic and overly sensitive to perturbation type, feature perturbation order, or categorical treatment (Barr et al., 2023).
• Normalization and Model-Comparability: Raw AOPC and other curve-based metrics are model-specific and can misleadingly favor certain architectures simply due to capacity and output range effects. Normalized AOPC (NAOPC) rescales scores by the empirically achievable lower and upper bounds, enabling fair comparison across models and inputs (Edin et al., 2024).
• Criticisms and Pathologies: Iterative masking-based metrics can conflate prediction robustness with faithfulness, especially when masking pushes examples out of distribution, resulting in unreliably high/low scores and large inter-model variance. Small perturbations or in-distribution transformations are preferred (Crothers et al., 2023).

A table summarizing select metric properties:

Metric	Target Domain	Faithfulness mechanism
Comprehensiveness/Sufficiency	Feature attribution (NLP/vision/tabular)	Output drop on (removal/retention) of important features
AOPC/NAOPC	Feature attribution (NLP/vision)	Area under perturbation curve, normalized per-model
PGI/PGI²	Tabular, trees	Expected output shift under perturbing top-k features
c-Eval	Vision/tabular	Minimal perturbation (non-explanatory features) to flip output
Adversarial Sensitivity	NLP	Distance between explanations under adversarial flip
LLM perturbation-based (LOO)	LLM QA/RAG	Sufficient span/keyword discovery; self-explanation overlap

5. Practical Guidelines and Limitations

Perturbation-based faithfulness metrics must be carefully selected and parameterized:

• Perturbation regime and context dependence: Hyperparameters (perturbation size, k/features ablated, masking operator) strongly affect conclusions; recommendations include matching perturbation strength to data scale, balancing granularity vs. cost (Barr et al., 2023, Gajewski et al., 2024, Fragkathoulas et al., 2024).
• Multiple metrics/reporting: Consensus is to report at least two faithfulness metrics, include sensitivity analyses (noise models, categorical handling), and cross-check with diagnosticity benchmarks (random attributions) (Barr et al., 2023, Chan et al., 2022).
• In-distribution perturbations: To avoid OOD pathologies, in-distribution transformations (e.g., adversarial attacks constrained by semantic similarity), and hybrid correctness metrics for evaluation are preferred (Manna et al., 2024, Fragkathoulas et al., 2024).
• Computational tractability: Techniques such as beam search NAOPC (Edin et al., 2024) and analytic PGI² for decision tree models (Gajewski et al., 2024) allow scaling to large datasets and complex models.

6. Emerging Directions and Theoretical Insights

Recent work advocates:

• Faithfulness with certified robustness: For structured models (e.g., Vision Transformers), faithfulness of attention-based explanations can be certified via randomized smoothing and denoised diffusion, with theoretical bounds on output/attention stability under Gaussian perturbations (Hu et al., 2023).
• Statistical significance frameworks: In LLMs, Distribution-Based Perturbation Analysis (DBPA) reframes perturbation-induced shifts as testable hypotheses, controlling for model stochasticity and yielding interpretable $p$-values for explanation impact (Rauba et al., 2024).
• Concept-level intervention: Closed-form concept perturbations (GRAD/ABL) yield reliable, automatic quantification of concept–output causal coupling, suggesting a path to generalizable, task-agnostic faithfulness evaluation (Li et al., 2024).

7. Contextualization within the Explainability Literature

Despite their centrality, perturbation-based faithfulness metrics have no universally endorsed variant; disagreements arise due to the diversity of applicable domains, model architectures, and perturbation semantics. Important caveats:

• Metrics must be explicitly matched to their intended domain of intervention and underlying model structure.
• Sensitivity to paradigmatic choices (e.g., masking vs. continuous noise) and hyperparameters remains significant and often underreported.
• Model-specific capacity and feature interaction patterns can invalidate cross-model comparability unless normalized metrics (e.g., NAOPC) are used.
• Faithfulness, as measured by perturbation, is at best a necessary but not sufficient condition for explanation utility, especially as complexity and distributional shifts increase.

Research continues on designing metrics that are robust, computationally efficient, and maximally aligned with both theoretical properties (e.g., monotonicity or informativeness given “ground truth” attributions) and human expectations for model transparency. The curation and careful application of perturbation-based faithfulness metrics remain a core methodological concern within the study of explainability and trustworthy artificial intelligence.