hhunugryy/aigc-security-iddm-anime-experiment

AIGC Security Experiment Materials

This repository contains public report materials for a course experiment on AIGC synthetic image detection and black-box evasion attacks.

The experiment uses an IDDM diffusion model trained on anime face images to generate synthetic images, evaluates a real-vs-synthetic detector, and compares traditional post-processing attacks with black-box pixel-level attacks.

Public Release Scope

Included:

• 1000 IDDM-generated anime face images.
• CEM black-box adversarial examples and their original generated-image counterparts.
• Experiment scripts.
• Experiment summaries, CSV/JSON result tables, and logs.
• Documentation for the generation process, detection setup, PPO exploration, and CEM attack.

Not included:

• Original real anime training images.
• Third-party detector weights.
• IDDM model checkpoints.
• The large post-train checkpoint directory.

Directory Structure

docs/
scripts/
results/
images/
  generated_iddm_1000/
  cem_original/
  cem_adv/
MANIFEST.txt

AIGC Image Generation

The synthetic images were generated by an IDDM diffusion model named animate_face_bonus.

Key generation settings:

model type: IDDM diffusion model
network: U-Net
condition mode: class conditional generation
class: class0
checkpoint: animate_face_bonus/ckpt_last.pt
sampling method: dpmpp
EMA weights: enabled
image size: 64 x 64
format: PNG
number of generated images: 1000

See:

docs/AIGC图像生成过程说明.md
docs/animate_face_bonus_model_report.md
images/generated_iddm_1000/
results/generated_iddm_1000_manifest.csv

Main Experimental Results

Baseline Detection

The detector achieved strong performance on real anime images and original IDDM-generated images:

AUC: 0.9976
IDDM generated images detected as synthetic: 995 / 1000

See:

results/exp01_baseline_detection/metrics.md

Traditional Attacks

JPEG compression and Gaussian noise strongly reduced the detector score.

JPEG q80 ASR: 0.9980
JPEG q80 SSIM: 0.9673

See:

results/exp02_traditional_attacks/attack_summary.md
results/exp04_quality_evaluation/quality_attack_summary.md

JPEG q80 Post-train

Using JPEG-q80 processed generated images for IDDM post-train did not transfer the evasion effect into the generator.

Post-train generated-image ASR: 0.0000
Mean detector score increased to: 2.3084

See:

results/exp03_posttrain_jpeg_q80/metrics.md
results/exp03_checkpoint_sweep/checkpoint_sweep_summary.md

PPO Black-box Attack

PPO was used as a learning-based black-box pixel perturbation strategy. It could reduce detector scores but had low evasion success.

PPO low-positive ASR: 0.0200

See:

results/exp03_ppo_default/
results/exp03_ppo_low_positive_eval_old/
results/exp03_ppo_low_positive_train/

CEM Black-box Attack

Cross-Entropy Method was used for per-image low-dimensional perturbation search. It does not use detector gradients, detector weights, JPEG compression, or white-box access.

CEM low-positive ASR: 0.4900
Mean score drop: 0.3799
Mean SSIM: 0.9933
Mean L_inf: 0.0380

See:

results/exp03_cem_low_positive/
images/cem_original/
images/cem_adv/

Intended Use

This release is intended for:

• Course report verification.
• AIGC detection robustness analysis.
• Reproducing result tables and qualitative comparisons.
• Studying black-box detector-score-based attack behavior.

Ethical and Safety Note

The black-box attack code and adversarial examples are provided only for academic security research and robustness evaluation. They should not be used to bypass production content moderation or provenance systems.

License and Data Note

This is a research-only release. The original real anime training images are not redistributed. Generated and adversarial images are included only as experiment artifacts derived from the trained generator and black-box attack pipeline.