Feature Attribution Stability Suite: How Stable Are Post-Hoc Attributions?

Kamalasankari Subramaniakuppusamy; Jugal Gajjar

cs.CVApr 2, 2026

Feature Attribution Stability Suite: How Stable Are Post-Hoc Attributions?

Authors: Kamalasankari Subramaniakuppusamy, Jugal Gajjar

Organizations: The George Washington University Washington D.C.

Abstract

Post-hoc feature attribution methods are widely deployed in safety-critical vision systems, yet their stability under realistic input perturbations remains poorly characterized. Existing metrics evaluate explanations primarily under additive noise, collapse stability to a single scalar, and fail to condition on prediction preservation, conflating explanation fragility with model sensitivity. We introduce the Feature Attribution Stability Suite (FASS), a benchmark that enforces prediction-invariance filtering, decomposes stability into three complementary metrics: structural similarity, rank correlation, and top-k Jaccard overlap-and evaluates across geometric, photometric, and compression perturbations. Evaluating four attribution methods (Integrated Gradients, GradientSHAP, Grad-CAM, LIME) across four architectures and three datasets-ImageNet-1K, MS COCO, and CIFAR-10, FASS shows that stability estimates depend critically on perturbation family and prediction-invariance filtering. Geometric perturbations expose substantially greater attribution instability than photometric changes, and without conditioning on prediction preservation, up to 99% of evaluated pairs involve changed predictions. Under this controlled evaluation, we observe consistent method-level trends, with Grad-CAM achieving the highest stability across datasets.

Figures & tables

Method	Pred. Inv.	Stab. Decomp.	$\epsilon$ -Noise	Geometric	Photometric	Compression
Local Lipschitz [ 3 ]	$\times$	$\times$	$\checkmark$	$\times$	$\times$	$\times$
Max-Sensitivity [ 34 ]	$\times$	$\times$	$\checkmark$	$\times$	$\times$	$\times$
OpenXAI [ 2 ]	$\times$	$\times$	$\checkmark$	$\times$	$\times$	$\times$
Quantus † [ 11 ]	$\checkmark$	$\times$	$\checkmark$	$\times$	$\times$	$\times$
LATEC [ 16 ]	$\times$	$\times$	$\checkmark$	$\times$	$\times$	$\times$
FASS (Ours)	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$

Table 1: Comparison of attribution stability evaluation approaches. Pred. Inv. : prediction invariance. Stab. Decomp. : stability decomposed into complementary failure-mode metrics (spatial, rank, salient-region). † Quantus assumes approximate prediction preservation but does not explicitly enforce prediction invariance through filtering, thereby potentially thereby failing to disentangle prediction changes from explanation instability.

Figure 1: FASS evaluation pipeline. Each input image is paired with its perturbed counterpart. Only prediction-invariant pairs proceed to attribution computation; excluded pairs are reported as retention diagnostics. Stability is decomposed into spatial (SSIM), ordinal (Spearman), and salient-region (Jaccard) components, averaged into a composite FASS score.

Perturbation	Mean	Std	Range
Rotation	30.9	27.5	0.0 – 88.1
Translation	0.1	0.2	0.0 – 0.6
Brightness	0.8	2.6	0.0 – 9.0
Noise	34.5	38.7	0.0 – 94.4
JPEG	1.0	3.4	0.0 – 11.7

Table 2: Prediction-invariant retention rates (%) by perturbation type (mean

\pm

std across all models and datasets).

Dataset	Mean	Std
CIFAR-10	11.5	15.3
ImageNet	63.3	6.2
COCO	28.9	37.7

Table 3: Prediction-invariant retention rates (%) by dataset (averaged across models and perturbations).

Dataset	Method	SSIM	Spn.	Jac.	FASS
ImageNet	IG	.706	.603	.060	.457
	GradSHAP	.681	.570	.037	.429
	Grad-CAM	.885	.966	.314	.722
	LIME	.342	.582	.072	.332
CIFAR-10	IG	.799	.700	.121	.540
	GradSHAP	.762	.643	.078	.494

Table 4: Mean FASS component scores per dataset, averaged across all models, perturbations, and retained prediction-invariant pairs. Spn. = Spearman; Jac. = Jaccard. Best FASS per dataset in bold .

Panel of Figure 2 — Figure 2: Attribution stability barplots across all three datasets. Top: ImageNet (native resolution; highest overall stability). Middle: CIFAR-10 (distribution-mismatch stress test; $32\times 32$ inputs upsampled $7\times$ ). Bottom: COCO (multi-object scenes; LIME narrows the gap with gradient-based methods). In all three settings Grad-CAM achieves the highest per-method FASS, and the method ranking Grad-CAM $>$ IG $>$ GradientSHAP $>$ LIME is preserved without rank crossings across all datasets.

Category	SSIM	Spn.	Jac.	FASS
Geometric	.725	.666	.099	.497
Photometric	.770	.724	.178	.557
Compression (JPEG)	.791	.739	.196	.576

Table 5: Mean component scores by perturbation category, averaged across all models, datasets, and methods (IG, GradientSHAP, Grad-CAM).

Appendix figures & tables27 assets

Supplementary material from the paper’s appendix.

Appendix

Architecture	Rot.	Trans.	Bright.	Noise	JPEG
ResNet-50	37.2	0.6	9.0	26.9	$<$ 0.1
DenseNet-121	20.8	$<$ 0.1	$<$ 0.1	12.1	$<$ 0.1
ConvNeXt-T	45.3	$<$ 0.1	$<$ 0.1	11.6	$<$ 0.1
ViT-B/16	36.1	$<$ 0.1	$<$ 0.1	29.2	$<$ 0.1

Appendix

Table 1: Prediction-invariant retention (%) on CIFAR-10.

Architecture	Rot.	Trans. †	Bright. †	Noise	JPEG †
ResNet-50	0.422	0.560	0.470	0.381	0.528
DenseNet-121	0.446	0.532	0.511	0.380	0.599
ConvNeXt-T	0.456	0.581	0.607	0.439	0.652
ViT-B/16	0.440	0.542	0.569	0.470	0.650

Appendix

Table 2: IG stability on CIFAR-10. † Near-zero retention; interpret with caution.

Architecture	Rot.	Trans. †	Bright. †	Noise	JPEG †
ResNet-50	0.423	0.501	0.425	0.393	0.468
DenseNet-121	0.445	0.491	0.479	0.384	0.520
ConvNeXt-T	0.457	0.536	0.563	0.441	0.583
ViT-B/16	0.442	0.517	0.523	0.463	0.588

Appendix

Table 3: GradientSHAP stability on CIFAR-10. † Near-zero retention.

Figure 2: GradientSHAP stability on CIFAR-10.

Architecture	Rot.	Trans. †	Bright. †	Noise	JPEG †
ResNet-50	0.450	0.529	0.643	0.449	0.703
DenseNet-121	0.535	0.523	0.735	0.553	0.760
ConvNeXt-T	0.548	0.834	0.796	0.645	0.774
ViT-B/16	0.539	0.667	0.744	0.598	0.722

Appendix

Table 4: Grad-CAM stability on CIFAR-10. † Near-zero retention.

Figure 3: Grad-CAM stability on CIFAR-10.

Architecture	Rot.	Trans. †	Bright. †	Noise	JPEG †
ResNet-50	0.282	0.346	0.346	0.284	0.350
DenseNet-121	0.279	0.339	0.338	0.273	0.344
ConvNeXt-T	0.295	0.362	0.356	0.288	0.364
ViT-B/16	0.280	0.368	0.357	0.287	0.437

Appendix

Table 5: LIME stability on CIFAR-10. † Near-zero retention.

Architecture	Rot.	Trans.	Bright.	Noise	JPEG
ResNet-50	58.8	$<$ 0.1	$<$ 0.1	70.1	$<$ 0.1
DenseNet-121	58.4	$<$ 0.1	$<$ 0.1	70.5	$<$ 0.1
ConvNeXt-T	68.3	$<$ 0.1	$<$ 0.1	71.3	$<$ 0.1
ViT-B/16	35.7	$<$ 0.1	$<$ 0.1	73.0	$<$ 0.1

Appendix

Table 6: Prediction-invariant retention (%) on ImageNet.

Architecture	Rot.	Trans. †	Bright. †	Noise	JPEG †
ResNet-50	0.471	0.418	0.502	0.642	0.541
DenseNet-121	0.466	0.409	0.498	0.642	0.538
ConvNeXt-T	0.483	0.432	0.521	0.613	0.562
ViT-B/16	0.497	0.448	0.553	0.768	0.589

Appendix

Table 7: IG stability on ImageNet. † Near-zero retention.

Architecture	Rot.	Trans. †	Bright. †	Noise	JPEG †
ResNet-50	0.470	0.421	0.472	0.572	0.492
DenseNet-121	0.465	0.412	0.468	0.572	0.487
ConvNeXt-T	0.484	0.438	0.501	0.568	0.519
ViT-B/16	0.490	0.443	0.512	0.626	0.531

Appendix

Table 8: GradientSHAP stability on ImageNet. † Near-zero retention.

Figure 6: GradientSHAP stability on ImageNet.

Architecture	Rot.	Trans. †	Bright. †	Noise	JPEG †
ResNet-50	0.726	0.671	0.821	0.877	0.853
DenseNet-121	0.761	0.703	0.873	0.915	0.889
ConvNeXt-T	0.779	0.724	0.869	0.897	0.878
ViT-B/16	0.762	0.712	0.842	0.822	0.852

Appendix

Table 9: Grad-CAM stability on ImageNet. † Near-zero retention.

Figure 7: Grad-CAM stability on ImageNet.

Architecture	Rot.	Trans. †	Bright. †	Noise	JPEG †
ResNet-50	0.410	0.340	0.390	0.470	0.430
DenseNet-121	0.370	0.360	0.410	0.490	0.450
ConvNeXt-T	0.390	0.350	0.430	0.470	0.440
ViT-B/16	0.400	0.380	0.440	0.570	0.460

Appendix

Table 10: LIME stability on ImageNet. † Near-zero retention.

Architecture	Rot.	Trans.	Bright.	Noise	JPEG
ResNet-50	88.1	$<$ 0.1	$<$ 0.1	74.3	11.7
DenseNet-121	51.7	$<$ 0.1	$<$ 0.1	90.6	$<$ 0.1
ConvNeXt-T	46.1	$<$ 0.1	$<$ 0.1	62.8	$<$ 0.1
ViT-B/16	45.4	$<$ 0.1	$<$ 0.1	76.7	0.3

Appendix

Table 11: Prediction-invariant retention (%) on COCO.

Architecture	Rot.	Trans. †	Bright. †	Noise	JPEG
ResNet-50	0.446	0.358	0.472	0.606	0.405
DenseNet-121	0.440	0.340	0.516	0.585	0.377 †
ConvNeXt-T	0.474	0.390	0.548	0.597	0.416 †
ViT-B/16	0.481	0.405	0.558	0.755	0.529

Appendix

Table 12: IG stability on COCO. † Near-zero retention.

Architecture	Rot.	Trans. †	Bright. †	Noise	JPEG
ResNet-50	0.445	0.364	0.425	0.533	0.391
DenseNet-121	0.437	0.339	0.442	0.519	0.367 †
ConvNeXt-T	0.475	0.393	0.506	0.568	0.411 †
ViT-B/16	0.481	0.404	0.515	0.647	0.494

Appendix

Table 13: GradientSHAP stability on COCO. † Near-zero retention.

Figure 10: GradientSHAP stability on COCO.

Architecture	Rot.	Trans. †	Bright. †	Noise	JPEG
ResNet-50	0.645	0.609	0.676	0.825	0.662
DenseNet-121	0.661	0.595	0.824	0.829	0.707 †
ConvNeXt-T	0.692	0.545	0.806	0.838	0.582 †
ViT-B/16	0.752	0.716	0.637	0.849	0.658

Appendix

Table 14: Grad-CAM stability on COCO. † Near-zero retention.

Architecture	Rot.	Trans. †	Bright. †	Noise	JPEG
ResNet-50	0.427	0.489	0.529	0.415	0.512
DenseNet-121	0.427	0.466	0.526	0.432	0.495 †
ConvNeXt-T	0.448	0.492	0.556	0.422	0.476 †
ViT-B/16	0.429	0.445	0.517	0.413	0.489

Appendix

Table 15: LIME stability on COCO. † Near-zero retention.

Explainability Evaluation

Explainability evaluation tests whether AI explanations are faithful, stable, understandable, useful, and reliable across models and settings.

See all 321 papers

Position: Explanation Stability Is a Property of the Model Method Pair, Not the Model
Jul 18, 2026Kabilan Elangovan, Daniel TingGradient-Based Attribution Explanation Stability
Singapore Health Services and Singapore Eye Research Institute, Singapore.
Learn to Rank: Visual Attribution by Learning Importance Ranking
Apr 7, 2026David Schinagl, Christian Fruhwirth-Reisinger, Alexander Prutsch +2Feature Attribution Explainable Image Classification
Institute of Visual Computing (IVC), Graz University of Technology, Graz, Austria · Nomadic AI, USA · Amazon, USA
Training for Trustworthy Saliency Maps: Adversarial Training Meets Feature-Map Smoothing
Mar 7, 2026Dipkamal Bhusal, Md Tanvirul Alam, Nidhi RastogiGradient-Based Attribution Explanation Stability
Rochester Institute of Technology

Jul 18, 2026J/K move · Enter open · S save

Position: Explanation Stability Is a Property of the Model Method Pair, Not the Model

Kabilan Elangovan, Daniel Ting

Singapore Health Services and Singapore Eye Research Institute, Singapore.

Gradient-Based Attribution Explanation Stability Explainable Medical Image Analysis Explainability Evaluation

Open paper AskPDF arXiv

This position paper argues that claims about explanation stability are scientifically invalid without cross method validation. Just as statistical significance requires the test statistic to be specified, stability should either be evaluated across multiple attribution paradigms or explicitly scoped to the computational objective of a single method. In controlled chest X ray experiments, DenseNet201, ResNet50V2, and InceptionV3 achieved AUC values above 99%, yet their stability rankings reversed across attribution methods. LayerCAM ranked InceptionV3 as the most stable model, with an IoU of 0.777, whereas GradCAM++ favored DenseNet201 and reduced InceptionV3 stability score by 17.3%. These findings demonstrate that explanation stability is an emergent property of the model method pair rather than an intrinsic characteristic of the model alone. We therefore argue that explanation based claims should be validated across multiple attribution methods and that regulatory submissions should explicitly specify the attribution operators used to avoid creating illusory safety assurances.

Abstract

Explainability Evaluation

Explore similar work

Position: Explanation Stability Is a Property of the Model Method Pair, Not the Model

Learn to Rank: Visual Attribution by Learning Importance Ranking

Training for Trustworthy Saliency Maps: Adversarial Training Meets Feature-Map Smoothing