Research map / Label-efficient and robust learning / Recognition and long-tail
F-Fidelity: A Robust Framework for Faithfulness Evaluation of Explainable AI
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes F-Fidelity, an explanation-agnostic fine-tuning plus stochastic-removal framework for evaluating XAI faithfulness, validated with degraded-explainer ground-truth rankings across images, time series, and text.
Problem
Evaluating post-hoc explainers via removal-based metrics (Fidelity, MoRF/LeRF) suffers from out-of-distribution inputs, while ROAR's retraining per explainer causes information leakage and is costly; explanation size is also hard to choose without ground truth.
Approach
Fine-tune the classifier once on randomly masked inputs (explanation-agnostic), then evaluate explainers with a stochastic removal that caps removed elements at a fraction beta of the input so the perturbed inputs stay in the fine-tuned model's distribution; also theoretically shows the metric's change-points reveal the true explanation tier size.
Key results
- On CIFAR-100 with SG-SQ, F-Fidelity achieves macro MoRF-vs-GT correlation of -1.00, versus -0.68 for Fidelity and -0.66 for ROAR.
- On Tiny ImageNet with GradCAM, F-Fidelity reaches -1.00 macro correlation across metrics, while ROAR and Fidelity are inconsistent (e.g., ROAR +0.63 MoRF vs GT).
- On NLP (SST2, BoolQ), vanilla Fidelity already performs near-perfectly, suggesting OOD effects are less pronounced for text models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use F-Fidelity instead of ROAR | ResNet explanations with GradCAM · CIFAR-100, macro MoRF vs GT | Spearman rank correlation: 0.99 -> -0.60 | Helped |
| Use F-Fidelity instead of Fidelity | ResNet explanations with GradCAM · CIFAR-100, micro MoRF vs GT | Spearman rank correlation: 0.57 -> -0.71 | Helped |
| Use F-Fidelity instead of R-Fidelity | LSTM explanations · Boiler, macro LeRF vs GT | Spearman rank correlation: 0.53 -> 0.86 | Helped |
| Use F-Fidelity instead of R-Fidelity | Transformer explanations · SST2, micro MoRF vs GT | Spearman rank correlation: -0.97 -> -1.00 | Helped |
| Fine-tune with random masking | ResNet on CIFAR-100 · CIFAR-100, 5% masking | classification accuracy: 56.70 -> 73.12 | Helped |
| Fine-tune with random masking | ResNet on CIFAR-100 · CIFAR-100, 10% masking | classification accuracy: 41.81 -> 72.97 | Helped |
| Fine-tune with random masking | LSTM on Boiler · Boiler, 10% masking | classification accuracy: 61.53 -> 73.19 | Helped |
| Fine-tune with random masking | LSTM on SST2 · SST2, 10% masking | classification accuracy: 79.82 -> 79.24 | Hurt |
| Fine-tune with random masking | Transformer on SST2 · SST2, 5% masking | classification accuracy: 77.29 -> 81.08 | Helped |
| Use F-Fidelity instead of Fidelity | LSTM explanations · SST2, macro MoRF vs GT | Spearman rank correlation: -1.00 -> -1.00 | No clear effect |
Surprising findings
- Vanilla Fidelity achieves near-perfect correlations on NLP tasks, matching or nearly matching F-Fidelity.
- Fidelity sometimes ranks completely random explainers above ground-truth-aligned ones on image tasks.
Limitations
- Relies on a degraded-explainer proxy for ground truth ranking rather than true explanations.
- Gains are marginal on NLP tasks where vanilla Fidelity already performs well.
- Theoretical size-recovery result assumes Shapley-based explainers with discrete influence tiers.
Experimental setup
Datasets: CIFAR-100 Tiny ImageNet PAM Boiler SST2 BoolQ
Benchmarks: CIFAR-100 Tiny ImageNet PAM Boiler SST2 BoolQ
Metrics: Spearman rank correlation (macro and micro) micro rank
Baselines: Fidelity ROAR R-Fidelity
Builds on: ROAR R-Fidelity Fidelity (graph explanation evaluation) Integrated Gradients SmoothGrad
Closest papers by meaning
- Navigating the Maze of Explainable AI: A Systematic Approach to Evaluating Methods and Metrics (NeurIPS 2024 · similarity 0.88)
- Respect the model: Fine-grained and Robust Explanation with Sharing Ratio Decomposition (ICLR 2024 · similarity 0.86)
- Faithful Model Explanations through Energy-Constrained Conformal Counterfactuals (AAAI 2024 · similarity 0.86)
- Evaluate with the Inverse: Efficient Approximation of Latent Explanation Quality Distribution (AAAI 2025 · similarity 0.86)
- GNNX-BENCH: Unravelling the Utility of Perturbation-based GNN Explainers through In-depth Benchmarking (ICLR 2024 · similarity 0.85)
- Counterfactual Explanations for Misclassified Images: How Human and Machine Explanations Differ (Abstract Reprint) (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.