Robustness of AI-Generated Image Detection Against Localized Inpainting Attacks

This repository hosts the detector-ready datasets and mask packs used in the thesis:

Robustness of AI-Generated Image Detection Against Localized Inpainting Attacks
Oguz Akin, Saarland University, CISPA Helmholtz Center for Information Security (2025)

It provides standardized evaluation splits for six state-of-the-art AI-generated image (AIGI) detectors across watermarking, passive, and training-free paradigms, tested under LaMa and ZITS inpainting attacks.

Everything is packaged as .tar.xz archives to ensure reproducibility and easy transfer.

📂 Repository Structure

. ├─ detectors/ │ ├─ ufd_datasets.tar.xz │ ├─ dimd_datasets.tar.xz │ ├─ dire_datasets.tar.xz │ ├─ aeroblade_datasets.tar.xz │ ├─ stablesig_datasets.tar.xz │ └─ treering_datasets.tar.xz ├─ masks/ │ ├─ masks_stablesig.tar.xz │ └─ masks_treering_wm.tar.xz └─ checksums.sha256

detectors/ — per-detector dataset “views,” already resized/re-encoded into the formats expected by each model.
masks/ — random-rectangle and random-blob object masks (area-binned), used to generate inpainting attacks.
checksums.sha256 — SHA-256 integrity hashes for all archives.

🔎 Dataset Details

Detector Views

Each archive expands into the exact layout expected by that detector. All splits contain 200 images per split (e.g. LaMa Inpainted Rand-Blob on SEMI-TRUTHS real images).

Typical layout: baseline/ reals/ fakes/ [fakes_inpainted_lama/, fakes_inpainted_zits/]

robustness/ inpainted_lama/ randrect/ randblob_bins/bin{1..4}/ inpainted_zits/ randrect/ randblob_bins/bin{1..4}/ [reals_inpainted/]

Detector Input Handling

On disk: All datasets are stored with their preprocessed versions for each detector to match their original paper/training setup.

UFD → 224
(Resized + center-cropped to 224×224, CLIP normalization.)
DIMD → JPEG-256
(Resized to 256×256, with JPEG round-trip to mimic training distribution.)
DIRE → 256
(Resized to 256×256, matching the ADM ImageNet-256 diffusion prior.)
AEROBLADE / StableSig / Tree-Ring → 512
(All evaluated directly at 512×512 without JPEG compression.)

Why this split? To eliminate the effect of compression or size on classification, ensuring scientifically fair evaluation.

Mask Packs

masks_stablesig.tar.xz
masks_treering_wm.tar.xz

Contain random rectangle and random blob masks, binned by area ratio:

bin1_0-3 → 0–3% of image area
bin2_3-10 → 3–10%
bin3_10-25 → 10–25%
bin4_25-40 → 25–40%

Used with LaMa and ZITS to create controlled inpainting attacks.

📐 Metrics

Datasets are organized to support a fixed-threshold robustness evaluation.

Baseline AUC
Distinguish clean reals vs fakes. Threshold t* chosen via Youden’s J.
Robustness AUC
Distinguish clean vs inpainted.
ΔAUC = Baseline – Robustness
ASR_inpainted (primary):
% of inpainted reals classified as Real at baseline t*.
ASR_fake→real (secondary):
% of baseline-detected fakes that flip to Real after inpainting.

Watermarking detectors:

Thresholds fixed at t90 and t99 on clean watermarked images together with another threshold that is determined at baseline to reflect a real life setting.
ASR = % attacked watermarked images where watermark is not detected.
AUC(clean vs attacked) sanity check.

📦 Archive Sizes

detectors/aeroblade_datasets.tar.xz — 1.5 GB
detectors/dimd_datasets.tar.xz — 117 MB
detectors/dire_datasets.tar.xz — 468 MB
detectors/stablesig_datasets.tar.xz — 924 MB
detectors/treering_datasets.tar.xz — 1.6 GB
detectors/ufd_datasets.tar.xz — 442 MB
masks/masks_stablesig.tar.xz — 2.2 MB
masks/masks_treering_wm.tar.xz — 1.2 MB

⚙️ Usage

Download & Extract

from huggingface_hub import hf_hub_download
import tarfile, os

REPO = "eoguzakin/Robustness of AI-Generated Image Detection Against Localized Inpainting Attacks"

def fetch_and_extract(filename, target_dir):
    path = hf_hub_download(repo_id=REPO, filename=filename, repo_type="dataset")
    os.makedirs(target_dir, exist_ok=True)
    with tarfile.open(path, "r:xz") as tar:
        tar.extractall(target_dir)
    print("Extracted:", target_dir)

# Example: UFD view + StableSig masks
fetch_and_extract("detectors/ufd_datasets.tar.xz", "/tmp/ufd")
fetch_and_extract("masks/masks_stablesig.tar.xz", "/tmp/masks_stablesig")

Integrity check

sha256sum -c checksums.sha256

🧪 Provenance Reals: SEMI-TRUTHS (Pal et al. 2024), OpenImages subset.

Fakes: GenImage diverse generator set.

Inpainting attacks: LaMa (Suvorov et al. 2022), ZITS (Dong et al. 2022).

Watermarks: Stable Signature (Fernandez et al. 2023), Tree-Ring (Wen et al. 2023).

Detector-specific preprocessing applied before runtime, ensuring comparability.

📸 Sample Images

Baseline (Real vs Fake)

Real	Fake

Inpainted reals (LaMa, ZITS & SEMI-TRUTHS)

LaMa	ZITS	SEMI-TRUTHS

LaMa	ZITS

Watermarks (StableSig vs Tree-Ring)

StableSig	LaMa	ZITS

Tree-Ring	LaMa	ZITS

📚 Citations If you use this dataset, please cite:

Pal et al., 2024 — Semi-Truths: A Large-Scale Dataset of AI-Augmented Images for Evaluating Robustness of AI-Generated Image Detectors.
Ojha et al., 2023 — Universal Fake Image Detectors.
Corvi et al., 2023 — On the Detection of Synthetic Images Generated by Diffusion Models.
Wang et al., 2023 — DIRE for Diffusion-Generated Image Detection.
Ricker et al., 2024 — AEROBLADE. Fernandez et al., 2023 — Stable Signature.
Wen et al., 2023 — Tree-Ring Watermarks.
Suvorov et al., 2022 — LaMa Inpainting.
Rombach et al., 2022 — Latent Diffusion Models. Dong et al., 2022 — ZITS Inpainting.

📝 License Derived datasets for research use only. Upstream datasets (SEMI-TRUTHS, GenImage, LaMa, ZITS, etc.) retain their original licenses. This packaging (scripts + archive structure) is released under CC BY-NC 4.0 unless otherwise specified.

👤 Maintainer Oguz Akin — Saarland University Contact: [email protected]

🗓️ Changelog v1.0 — Initial release with detector views + masks for LaMa and ZITS inpainting.

image imagewidth (px) 512 512