ashleyscruse/noble-ai-evidence-benchmark

NOBLE AI-Generated Evidence Detection Benchmark

A domain-specific benchmark for evaluating AI-generated image detection tools on law enforcement imagery (surveillance footage, bodycam, evidence-style photos). The benchmark spans multiple generator architectures and three image-quality levels designed to mimic the conditions in which real evidence reaches courtrooms.

Status: v1.1 release. All three generators (FLUX-schnell, Realistic Vision 5.1, SDXL) complete, paired-prompt design realized across the full benchmark.

Why this dataset exists

Existing AI image detection tools were trained on social media imagery: clean, well-lit, high-resolution. The reality of law enforcement footage is none of those things. Bodycam, dashcam, and surveillance video are typically grainy, compressed, and recorded under variable lighting. Detection tools that work on social media content can fail dramatically on the very imagery courts must evaluate when AI-generated material is presented as evidence.

This benchmark fills that gap. It is the first dataset designed specifically to test AI-image detectors on degraded, domain-matched law enforcement content.

Dataset composition

Real images

Source	Count	Notes
UCF Crime surveillance frames	7,746	Stratified sample across 14 anomaly + normal categories
Total real	7,746

Synthetic images

Generator	Count	Architecture	Model ID
Realistic Vision 5.1	3,600	SD 1.5 fine-tune (UNet)	SG161222/Realistic_Vision_V5.1_noVAE
SDXL	3,600	UNet, multi-stage	stabilityai/stable-diffusion-xl-base-1.0
FLUX.1-schnell	3,600	DiT (transformer)	black-forest-labs/FLUX.1-schnell
Total synthetic	10,800

Each generator used the same 240 prompts (paired-prompt design), 15 variations per prompt, with matched seeds across all three generators. This enables paired statistical comparison.

Degradation levels

Level	Parameters	Simulates
Clean	None	High-quality digital photos
Moderate	JPEG Q50, blur sigma 1.0, contrast 0.8x	Decent surveillance footage
Heavy	JPEG Q30, downscale 0.5x, noise sigma 25, blur sigma 2.0	Poor bodycam, old CCTV

Dataset size summary

Split	Count
Raw real	7,746
Raw synthetic	10,800
Processed real (3 levels)	23,238
Processed synthetic (3 levels x 3 generators)	32,400
Total image instances	74,184

Dataset structure

The dataset is organized by split, degradation level, and source. See manifest.csv for a complete file index with metadata columns: path, label, split, level, category, generator, prompt_id, variation_id.

File naming convention for synthetic images: model_pPROMPTID_vVARIATIONID.png (e.g., rv51_p0042_v003.png), enabling matched-pair comparison across generators.

Intended uses

• Benchmarking AI-generated image detection tools on domain-specific content
• Studying the effect of image degradation on detection accuracy
• Studying generalization gaps across generator architectures (UNet vs. DiT)
• Training and fine-tuning improved detection models for law enforcement contexts
• Informing policy discussions on AI evidence admissibility (e.g., Federal Rule 707)

Out-of-scope uses

• Real-time evidence authentication without further calibration on operational data
• Standalone forensic verification of disputed evidence in legal proceedings
• Training generative models intended to fabricate law enforcement imagery

Licensing

The benchmark structure, organization, metadata, and associated code are released under CC-BY-NC-4.0.

This dataset is compositional. Downstream users must respect source-material licenses:

• UCF Crime videos: research-use license (Sultani et al., 2018)
• Realistic Vision 5.1 outputs: CreativeML Open RAIL-M
• SDXL outputs: CreativeML Open RAIL++
• FLUX.1-schnell outputs: Apache 2.0

Citation

@dataset{scruse2026noble,
  author    = {Scruse, Ashley},
  title     = {NOBLE AI-Generated Evidence Detection Benchmark},
  year      = {2026},
  publisher = {HuggingFace},
  version   = {1.1}
}

Funding

National Organization of Black Law Enforcement Executives (NOBLE) research grant to Morehouse College. Compute via TACC Vista and Stampede3 under NSF ACCESS allocation TRA25001.

Contact

Ashley Scruse, PhD -- ashley.scruse@morehouse.edu Postdoctoral Researcher, Center for Broadening Participation in Computing, Morehouse College

Version history

• v1.1 (2026-05-11): Full release. Three generators (FLUX, RV51, SDXL), three degradation levels, 74,184 total instances.
• v1.0 (2026-05-06): Initial release with RV51 + SDXL only.