Deepfake Detector Robustness Testing

danbabalola2 pts1 comments

danb21/social-media-robustness-sdxl-instantid · Datasets at Hugging Face

Log In<br>Sign Up

An evaluation benchmark released under CC BY-NC 4.0 for research evaluation only, not for training detection models or commercial use. Optional: tell me what you work on, and opt in below if you want a heads-up when datasets like this drop. I plan the next dataset around what people actually need.\n","classNames":"hf-sanitized hf-sanitized-Ab7YwAptoZL5_q4eSzBqU"},"customHeading":"Request access to the Social Media Robustness Benchmark","gated":"auto","isLoggedIn":false,"repoId":"danb21/social-media-robustness-sdxl-instantid","repoType":"dataset","requiresPaidPlan":false}"> Request access to the Social Media Robustness Benchmark<br>This repository is publicly accessible, but you have to accept the conditions to access its files and content.<br>An evaluation benchmark released under CC BY-NC 4.0 for research evaluation only, not for training detection models or commercial use. Optional: tell me what you work on, and opt in below if you want a heads-up when datasets like this drop. I plan the next dataset around what people actually need.

Log in or Sign Up to review the conditions and access this dataset content.

Social Media Robustness Benchmark: SDXL+InstantID Synthetic Face Detection

Version: v1.0.0 · License: CC BY-NC 4.0 (research evaluation only)

Detector accuracy on clean lab test sets does not predict in-the-wild performance. Social<br>platforms re-encode every uploaded image: platform-specific JPEG, resize, chroma subsampling,<br>metadata stripped. This benchmark lets detector authors and procurers measure robustness under<br>documented, paired, demographically-balanced conditions instead of blunt lab proxies.

A companion blog post and white paper cover the methodology, statistics, and findings in full.<br>This card describes what the dataset is, how it is built at a high level, and how to use it.

1. Details

Field<br>Value

Name<br>Social Media Robustness Benchmark: SDXL+InstantID Synthetic Face Detection

Version<br>v1.0.0

Base corpus<br>2,400 images (1,200 real, 1,200 generated), sampled from danb21/synthetic-face-sdxl-instantid-bench at a deterministic per-cell quota

Perturbations<br>12 single-axis lab variants (JPEG, resize, noise, blur) + 4 platform-pipeline approximations (Instagram, Facebook, TikTok, X)

Total rows<br>40,574 image rows across 17 configurations (2,400 clean + 28,800 lab + 9,374 platform)

Cell axis<br>6 skin tones × 2 genders × 2 labels, 100 per cell at the clean baseline

Pairing<br>Every configuration evaluates the same images (paired by media_id)

Blog post<br>Read the write-up (companion post; published before the paper)

Paper<br>In preparation; methodology and results reported there

Maintainer<br>Daniel Babalola, danielbabalola@alumni.upenn.edu

2. What This Dataset Is For

Use it to:

Compute paired AUC deltas, AUC(clean) − AUC(perturbation), per detector per condition.

Measure per-cell robustness (skin tone × gender) under each perturbation.

Compare detector architectures under matched conditions.

It is not training data . It is small by design (2,400 base images), paired by construction<br>(every condition evaluates the same images), and the platform pipelines are calibrated<br>approximations of each platform's mean re-encode behaviour, not pixel-faithful platform<br>reproductions (see Limitations).

3. Structure

Configurations

Config<br>Layer<br>Description

clean<br>clean<br>2,400 unperturbed base images; the reference for every paired delta

layer1_jpeg_q{30,50,70,80,95}<br>lab<br>JPEG re-encode at the named quality factor

layer1_resize_{0.5,0.75}<br>lab<br>Bicubic downsample then upsample back

layer1_noise_{5,10}<br>lab<br>Additive Gaussian noise (variance 5 / 10)

layer1_blur_{1,2,4}<br>lab<br>Gaussian blur (sigma 1 / 2 / 4)

layer2_ig_pipeline<br>platform<br>Instagram (JPEG ~92, max edge 1440, 4:2:0, EXIF stripped)

layer2_fb_pipeline<br>platform<br>Facebook (JPEG ~93, max edge 1920, 4:2:0, EXIF stripped)

layer2_tt_pipeline<br>platform<br>TikTok (JPEG ~80, max edge 1920, 4:2:0)

layer2_x_pipeline<br>platform<br>X (JPEG ~93, max edge 1920, 4:2:0, EXIF stripped)

All configurations share the same column schema. Key columns: image, label<br>(real/generated), cell_skin_tone, cell_gender, media_id (stable across<br>configurations for pairing), perturbation_slug, perturbation_layer, and the<br>measured-encoding fields.

Balance

The clean baseline and all 12 lab configurations are uniform at 100 images per cell per<br>label (6 skin tones × 2 genders × real/generated). The 4 platform configurations re-crop the<br>laundered output back to a 256×256 face crop for comparability; all real-side cells retain<br>100/100, while a small fraction of synthetic-face crops do not survive re-detection (~2,342–2,344<br>rows per platform config). That concentration is itself a finding, analyzed in the white paper.

4. How It Is Built (high level)

Source. A deterministic per-cell subset of the v1 Synthetic Face Detection Benchmark:<br>1,200 real Pexels frames and 1,200 SDXL+InstantID outputs, uniform at 100 per cell per...

platform robustness benchmark social face clean

Related Articles