Forgery Localization Eval

forgery-localization-eval

Detecting AI-Generated Forgeries via Iterative Manifold Deviation Amplification — Jiangling Zhang et al. (2026) (arXiv:2602.18842, 2026)

What this evaluates

Evaluates pixel-level localization accuracy for detecting AI-generated and traditionally tampered image forgeries. It probes a model's ability to distinguish manipulated regions from authentic content by measuring spatial overlap and detection trade-offs against ground-truth masks.

Datasets

• OpenSDID — total ?; splits: train (-1), test (-1)
• GIT10K — total 10000; splits: test (-1)
• CocoGlide — total ?; splits: test (-1)
• Inpaint32K — total ?; splits: test (-1)
• IMD2020 — total 2010; splits: test (-1)
• NIST16 — total ?; splits: test (-1)
• CASIA — total ?; splits: test (-1)

Metrics

• F1-score (primary) — range: [0, 1]
  • Harmonic mean of precision and recall: 2 * (precision * recall) / (precision + recall). Reflects the trade-off between detection precision and recall.
• IoU — range: [0, 1]
  • Intersection over Union: |predicted_mask ∩ ground_truth_mask| / |predicted_mask ∪ ground_truth_mask|. Quantifies spatial overlap between predicted and ground-truth masks.

Input / output format

Input: RGB images resized to 512×512 pixels. The model processes a dual-stream input combining the original image and MAE reconstruction residuals.

Output: Pixel-level binary/soft masks indicating the spatial location of manipulated/forged regions.

Scoring recipe

def compute_metrics(pred_mask, gt_mask):
    pred = (pred_mask > 0.5).astype(bool)
    gt = (gt_mask > 0.5).astype(bool)
    intersection = np.logical_and(pred, gt).sum()
    union = np.logical_or(pred, gt).sum()
    iou = intersection / union if union > 0 else 0.0
    tp = intersection
    fp = pred.sum() - tp
    fn = gt.sum() - tp
    precision = tp / (tp + fp) if (tp + fp) > 0 else 0.0
    recall = tp / (tp + fn) if (tp + fn) > 0 else 0.0
    f1 = 2 * precision * recall / (precision + recall) if (precision + recall) > 0 else 0.0
    return iou, f1

Common pitfalls

• Models are pretrained on OpenSDID and fine-tuned per dataset, so reported scores reflect fine-tuned performance rather than zero-shot generalization.
• All inputs are resized to 512×512, which may artificially inflate or deflate IoU/F1 on high-resolution benchmarks like NIST16 compared to native-resolution evaluation.
• Metrics are averaged across datasets in the paper (GIT-AVG, TT-AVG), but individual dataset performance varies significantly due to differing manipulation types and mask complexities.

Evidence (verbatim from paper)

Following common practice, we report F1-score and Intersection-over-Union (IoU). F1 reflects the trade-off between precision and recall, indicating the overall detection capability, while IoU quantifies the spatial overlap between the predicted and ground-truth masks, highlighting localization accuracy.

Citation

@misc{zhang2026detecting,
  title={Detecting AI-Generated Forgeries via Iterative Manifold Deviation Amplification},
  author={Jiangling Zhang et al. (2026)},
  year={2026},
  note={arXiv:2602.18842}
}

• arXiv: 2602.18842