VAAS: Vision-Attention Anomaly Scoring for Image Manipulation Detection in Digital Forensics

Abstract

Recent advances in AI-driven image generation have introduced new challenges for verifying the authenticity of digital evidence in forensic investigations. Modern generative models can produce visually consistent forgeries that evade traditional detectors based on pixel or compression artefacts. Most existing approaches also lack an explicit measure of anomaly intensity, which limits their ability to quantify the severity of manipulation. This paper introduces Vision-Attention Anomaly Scoring (VAAS), a novel dual-module framework that integrates global attention-based anomaly estimation using Vision Transformers (ViT) with patch-level self-consistency scoring derived from SegFormer embeddings. The hybrid formulation provides a continuous and interpretable anomaly score that reflects both the location and degree of manipulation. Evaluations on the DF2023 and CASIA v2.0 datasets demonstrate that VAAS achieves competitive F1 and IoU performance while enhancing visual explainability through attention-guided anomaly maps. The framework bridges quantitative detection with human-understandable reasoning, supporting transparent and reliable image integrity assessment.

This repository contains the experimental training and evaluation code used in the VAAS study.

Model Framework

Inference Sample Visuals

Read the research paper

• arXiv version
• [Conference version](to be added after publication)

Using model: Huggingface Model Card for Quick Inference and use cases

• VAAS-v1 on Huggingface
• See other VAAS model versions for usage and contribution

Dataset

CASIA v2.0

CASIA by Dong is a foundational benchmark for evaluating splicing and compositing detection methods. It contains approximately 12,614 images, including 7,491 authentic and 5,123 tampered samples, each with a corresponding ground-truth mask.

• Casia v2.0 dataset link

DF2023

DF2023 by Fischinger and Boyer was designed to benchmark forensic methods against AI-generated and hybrid manipulations. The full dataset contains approximately one million forged images distributed across four manipulation types: 100K removal, 200K enhancement, 300K copy–move, and 400K splicing operations.

• DF2023 dataset link

Dependencies

uv sync

• PyTorch with CUDA support recommended
• Python 3.10+

Training and Evaluating CASIA

  python train.py \
  --dataset CASIA2 \
  --dataset-root "<Absolute path to dataset root directory>/CASIA2/CASIA2" \
  --exp-id "CASIA_segformer_v3" \
  --epochs 50 \
  --lr 1e-4 \
  --loss-type focal \
  --pos-weight 15 \
  --dice-weight 0.7 \
  --focal-alpha 0.25 \
  --focal-gamma 2.5 \
  --alpha 0.3

Inference

  python infer.py \
  --dataset-root "<Absolute path to dataset root directory>/CASIA2/CASIA2" \
  --dataset CASIA2 \
  --checkpoint-dir "<Absolute path to checkpoint>" \
  --output-dir "inference_visuals" \
  --num-samples 20 \
  --vis-mode both \
  --alpha 0.5

Threshold sweep

  python threshold_sweep.py \
  --dataset-root "<Absolute path to dataset root directory>/CASIA2/CASIA2" \
  --checkpoint-dir "Absolute path to checkpoint>" \
  --max-samples 300

Training and Evaluating DF2023

  python train.py \
  --dataset DF2023 \
  --dataset-root "<Absolute path to dataset root directory>/DF2023_V15" \
  --exp-id "DF2023_segformer_v3" \
  --epochs 50 \
  --lr 1e-4 \
  --loss-type focal \
  --pos-weight 15 \
  --dice-weight 0.7 \
  --focal-alpha 0.25 \
  --focal-gamma 2.5 \
  --alpha 0.3

Inference

  python infer.py \
  --dataset-root "<Absolute path to dataset root directory>/DF2023_V15" \
  --dataset DF2023 \
  --checkpoint-dir "<Absolute path to checkpoint>" \
  --output-dir "inference_visuals_new" \
  --num-samples 20 \
  --vis-mode both \
  --alpha 0.5

Threshold sweep

  python threshold_sweep.py \
  --dataset-root "<Absolute path to dataset root directory>/DF2023_V15" \
  --checkpoint-dir "<Absolute path to checkpoint>" \
  --max-samples 300

Citation

If you use VAAS in your research, please cite both the paper and the associated software as appropriate.

@article{bamigbade2025vaas,
  title={VAAS: Vision-Attention Anomaly Scoring for Image Manipulation Detection in Digital Forensics},
  author={Bamigbade, Opeyemi and Scanlon, Mark and Sheppard, John},
  journal={arXiv preprint arXiv:2512.15512},
  year={2025}
}

@software{vaas,
  title        = {VAAS: Vision-Attention Anomaly Scoring},
  author       = {Bamigbade, Opeyemi and Scanlon, Mark and Sheppard, John},
  year         = {2025},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.18064355},
  url          = {https://doi.org/10.5281/zenodo.18064355}
}

---