Research map / Learning theory and algorithms / Prediction and decision losses
MFABA: A More Faithful and Accelerated Boundary-Based Attribution Method for Deep Neural Networks
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 MFABA, a boundary-based attribution method that uses adversarial gradient ascent with a second-order Taylor-derived trapezoid-style sum, claiming axiomatic validity and over 100x speedup over BIG and IG.
Problem
Existing attribution methods (IG, BIG, AGI, GIG) either lack faithfulness or are computationally expensive, requiring many forward and backward passes to compute attributions for DNNs.
Approach
Attribution is computed along an adversarial gradient-ascent path, summing averaged endpoint gradients times input increments (a second-order Taylor-based approximation using the Hessian replaced by gradient differences), reusing stored gradients so that only a few (3-10) forward and backward passes are needed; softmax outputs are used as the target and sharp/smooth gradient normalizations are off
Key results
- MFABA reports 136.9630 FPS vs 1.3492 for BIG on CIFAR10 ResNet-50 (over 101x faster), per the FPS table.
- Error rate on CIFAR10 ResNet-50 drops from 0.2613 (vanilla) to 0.01165 with MFABA.
- On ResNet-50 insertion score, MFABA(SHARP) reaches 0.3237 vs 0.3695 for AGI, so AGI is higher on that metric; MFABA leads on AUC (0.4857 vs 0.4823 for IG).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use second-order Taylor expansion instead of first-order only (ablation) | MFABA · ImageNet with EfficientNet | Error Rate: 0.02058 -> 0.01165 | Helped |
| Use second-order Taylor expansion instead of first-order only (ablation) | MFABA · CIFAR10 with ResNet-50 | Error Rate: 0.2613 -> 0.01165 | Helped |
| Use second-order Taylor expansion instead of first-order only (ablation) | MFABA · CIFAR100 with ResNet-50 | Error Rate: 0.10221 -> 0.04192 | Helped |
| Use MFABA instead of BIG | Attribution on Inception-v3 · Inception-v3 | Insertion score: 0.4840 -> 0.5407 | Helped |
| Use MFABA instead of AGI | Attribution on Inception-v3 · Inception-v3 | Deletion score: 0.0590 -> 0.0627 | Hurt |
| Use MFABA instead of BIG | Attribution on Inception-v3 · Inception-v3 | AUC: 0.5200 -> 0.5389 | Helped |
| Use MFABA instead of AGI | Attribution on ResNet-50 · ResNet-50 | Insertion score: 0.3695 -> 0.3237 | Hurt |
| Use MFABA instead of BIG | Attribution on ResNet-50 · ResNet-50 | Deletion score: 0.0485 -> 0.0566 | Hurt |
| Use MFABA instead of AGI | Attribution on ResNet-50 · ResNet-50 | AUC: 0.4772 -> 0.4857 | Helped |
| Use MFABA instead of AGI | Attribution on VGG16 · VGG16 | Insertion score: 0.2910 -> 0.2856 | Hurt |
| Use MFABA instead of BIG | Attribution on VGG16 · VGG16 | Deletion score: 0.0390 -> 0.0410 | Hurt |
| Use MFABA instead of AGI | Attribution on VGG16 · VGG16 | AUC: 0.4359 -> 0.4540 | Helped |
Surprising findings
- On ResNet-50 and VGG16 the deletion score of IG and AGI is lower than MFABA, so MFABA is not uniformly best on the insertion/deletion metrics.
- The error-rate table reports the same MFABA value (0.01165) for ImageNet EfficientNet and CIFAR10 ResNet-50, which looks like a possible reporting error.
Limitations
- Evaluation limited to conventional image datasets and classification models, as the authors acknowledge.
- Mixed results on insertion and deletion scores; AGI and IG beat MFABA on some metrics.
- Several theoretical steps (Taylor expansion, Hessian approximation) are heuristic and the paper omits the epsilon term, weakening the axiomatic claims.
Experimental setup
Datasets: CIFAR10 CIFAR100 ImageNet
Benchmarks: CIFAR10 CIFAR100 ImageNet
Metrics: error rate insertion score deletion score AUC FPS
Baselines: Integrated Gradients Saliency Map BIG AGI SmoothGrad DeepLift Vanilla MFABA
Builds on: Integrated Gradients Boundary-based Integrated Gradients Adversarial Gradient Integration FGSM I-FGSM
Closest papers by meaning
- Axiomatic Attribution for Deep Networks (ICML 2017 · similarity 0.89)
- One Wave To Explain Them All: A Unifying Perspective On Feature Attribution (ICML 2025 · similarity 0.85)
- Iterative Search Attribution for Deep Neural Networks (ICML 2024 · similarity 0.85)
- How to Probe: Simple Yet Effective Techniques for Improving Post-hoc Explanations (ICLR 2025 · similarity 0.84)
- Benchmarking the Attribution Quality of Vision Models (NeurIPS 2024 · similarity 0.84)
- quanda: An Interpretability Toolkit for Training Data Attribution Evaluation (NeurIPS 2026 · similarity 0.83)
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.