Deepfake detection, measured after the repost

The detectors work in the lab. This is where they break in the wild.

Every deepfake detector reports near-perfect accuracy. This is the benchmark that measures what happens after the file gets compressed, re-encoded, and reposted, across video, audio, and image. Open, reproducible, and honest about the gap.

Read this first. Reference run on synthetic fixtures. These numbers are real and reproducible (deeptruth reference), but the data is synthetic, generated to exercise the harness. They are not FaceForensics++/ASVspoof results. Real detectors are shown as pending until run on the real, gated datasets.

The gap table

The number a paper prints is the clean, in-distribution one. The number that matters is the one after the file has traveled. This is both, side by side. Larger gap, bigger surprise.

DetectorModalityStatusIn-distributionAfter it travelsThe gap
dct-linear
frequency · synthetic-images
imagereference95.4%64.4%
worst: 49.6% (rescale_8x)
-31.0%
frame-frequency
frequency · synthetic-video
videoreference100.0%49.9%
worst: 47.5% (h264_crf45)
-50.1%
radial-spectrum
frequency · synthetic-images
imagereference94.2%61.5%
worst: 49.2% (jpeg_q10)
-32.7%
spectral-audio
audio-antispoofing · synthetic-audio
audioreference100.0%81.3%
worst: 50.0% (mp3_32k)
-18.8%

Degradation curves

Accuracy does not fall off a cliff at one magic setting; it bleeds out across a whole range of ordinary insults. Here is each detector's AUC as compression, resolution loss, and mild perturbation get worse. Every point carries a bootstrap confidence interval; hover for exact values.

dct-linear

reference

frequency · synthetic-images · Frank et al., Leveraging Frequency Analysis for Deep Fake Recognition, ICML 2020

Clean AUC0.99395% CI 0.9860.998
Clean acc95.4%n = 240
ECE0.040calibration

AUC vs degradation severity. The dashed line at 0.50 is a coin flip: a curve that reaches it has no signal left.

instagram: 78.3%x: 92.1%whatsapp: 53.3%

frame-frequency

reference

frequency · synthetic-video · Frame-level radial spectrum (cf. Durall et al., CVPR 2020), mean-pooled

Clean AUC1.00095% CI 1.0001.000
Clean acc100.0%n = 80
ECE0.111calibration

AUC vs degradation severity. The dashed line at 0.50 is a coin flip: a curve that reaches it has no signal left.

instagram: 50.0%tiktok: 50.0%x: 50.0%

radial-spectrum

reference

frequency · synthetic-images · Durall et al., Watch your Up-Convolution, CVPR 2020

Clean AUC0.98395% CI 0.9690.994
Clean acc94.2%n = 240
ECE0.058calibration

AUC vs degradation severity. The dashed line at 0.50 is a coin flip: a curve that reaches it has no signal left.

instagram: 60.8%x: 83.8%whatsapp: 50.4%

spectral-audio

reference

audio-antispoofing · synthetic-audio · Spectral front-end baseline (cf. AASIST, Jung et al., ICASSP 2022)

Clean AUC1.00095% CI 1.0001.000
Clean acc100.0%n = 160
ECE0.035calibration

AUC vs degradation severity. The dashed line at 0.50 is a coin flip: a curve that reaches it has no signal left.

Pending: the real detectors

These are wired up and waiting on a GPU run against the real, gated datasets. They show here with no numbers, on purpose. A benchmark that invents results for models it never ran would be the exact dishonesty this project exists to call out.

aasist

audio · audio-antispoofing
Jung et al., AASIST, ICASSP 2022

clip-universalfakedetect

image · transformer
Ojha et al., Towards Universal Fake Image Detectors, CVPR 2023

efficientnet-b4

video · cnn
Bonettini et al., Video Face Manipulation Detection Through Ensemble of CNNs, ICPR 2020

xception-ffpp

video · cnn
Rossler et al., FaceForensics++, ICCV 2019