Spectral Integrated Gradients for Coarse-to-Fine Feature Attribution

This repository provides the source code for our paper "Spectral Integrated Gradients for Coarse-to-Fine Feature Attribution," accepted to ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) 2026.

SIG (Spectral Integrated Gradients) is a feature attribution method that constructs the integration path in the frequency domain, traversing from coarse (low-frequency) to fine (high-frequency) image components. This coarse-to-fine path yields sharper and more faithful attributions than path-integral baselines that interpolate directly in pixel space.

Installation

conda create -n sig python=3.9
conda activate sig
pip install -e .
pip install -r requirements.txt

The provided requirements.txt pins torch==2.7.0 (CUDA 12.8 wheel) and torchvision==0.22.0. Adjust the --extra-index-url line in requirements.txt if you need a different CUDA build.

Datasets

We evaluate on three image classification datasets (and CIFAR-10 for ROAR):

• ImageNet (ILSVRC2012) — standard validation split.
• Oxford-IIIT Pet — 37 fine-grained breeds.
• Oxford Flowers 102 — 102 flower species.
• CIFAR-10 — used for the ROAR benchmark.

Datasets are not shipped with this repository. Edit the dataset_path field in the corresponding YAML under configs/dataset/ before running any script:

# configs/dataset/oxfordpet.yaml
dataset_path: /path/to/oxfordpet/images

# configs/dataset/oxfordflower.yaml
dataset_path: /path/to/oxfordflower

# configs/dataset/imagenet.yaml
dataset_path: /path/to/imagenet2012

Pretrained Classifiers

We provide the fine-tuned classifier checkpoints used in the paper for Oxford-IIIT Pet and Oxford Flowers 102 under checkpoints/:

checkpoints/
├── classifier_oxfordpet/
│   ├── vgg16_best.pt
│   ├── resnet18_best.pt
│   └── inception_best.pt
└── classifier_oxfordflower/
    ├── vgg16_best.pt
    ├── resnet18_best.pt
    └── inception_best.pt

For ImageNet we use torchvision's pretrained weights directly, so no checkpoint is required. Checkpoint loading is handled in scripts/diffid.py, which resolves paths relative to the repository root.

DiffID Benchmark

Each attribution method has a runner script under scripts/benchmark_diffid/:

# Our method (SIG)
bash scripts/benchmark_diffid/spectral_ig.sh <overlap>

# Baseline methods
bash scripts/benchmark_diffid/ig.sh
bash scripts/benchmark_diffid/big.sh
bash scripts/benchmark_diffid/ig2.sh
bash scripts/benchmark_diffid/gig.sh
bash scripts/benchmark_diffid/agi.sh
bash scripts/benchmark_diffid/eig.sh
bash scripts/benchmark_diffid/mig.sh
bash scripts/benchmark_diffid/samp.sh
bash scripts/benchmark_diffid/grad_input.sh

Each script iterates over:

• Datasets: ImageNet, Oxford-IIIT Pet, Oxford Flowers 102
• Classifiers: VGG-16, ResNet-18, Inception

Results are written to results/benchmark_diffid/<dataset>/<method>/<model>/.

ROAR Benchmark

ROAR-style benchmarks (CIFAR-10) are provided under scripts/benchmark_roar/:

# Our method (SIG)
bash scripts/benchmark_roar/spectral_ig.sh

# Baseline methods
bash scripts/benchmark_roar/ig.sh
bash scripts/benchmark_roar/big.sh
bash scripts/benchmark_roar/ig2.sh
bash scripts/benchmark_roar/gig.sh
bash scripts/benchmark_roar/agi.sh
bash scripts/benchmark_roar/eig.sh
bash scripts/benchmark_roar/mig.sh
bash scripts/benchmark_roar/samp.sh
bash scripts/benchmark_roar/grad_input.sh
bash scripts/benchmark_roar/random.sh

Results are written to results/benchmark_roar/<dataset>/<method>/.

Citation

If you use this codebase, please consider citing:

@inproceedings{kim2026spectral,
  title={Spectral Integrated Gradients for Coarse-to-Fine Feature Attribution},
  author={Kim, Soyeon and Lim, Seongwoo and Lee, Kyowoon and Choi, Jaesik},
  booktitle={Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)},
  year={2026},
  url={http://arxiv.org/abs/2605.19607}
}

License

This project is released under the MIT License.