Deepfake detection on social media: what re-encoding does to detector scores

This page belongs to the deepfake detection hub and follows the platform path step by step. The compression and deepfake detection page gives the full compression tables; this one is about what happens to a clip between the generator and the feed, and what each step costs.

01What does a platform do to an uploaded video?

A video uploaded to a social platform is not stored as sent. The platform transcodes it to its own codec settings and bitrate ladder, often at several resolutions, and serves whichever rendition fits the viewer's connection. The FaceForensics++ authors built their three release levels to imitate exactly this: they wrote that raw videos "are rarely found on the internet" and compressed every clip with the H.264 codec, "which is widely used by social networks or video-sharing websites" (arXiv 1901.08971). Their HQ level uses a constant rate quantisation parameter of 23, which the paper calls visually nearly lossless, and their LQ level uses 40.

Transcoding is only the first step. A clip that spreads is cropped, captioned, screen recorded from one app and uploaded to another, and transcoded again at each hop, so the original encoding, resolution and framing are unknown by the time a detector sees it.

02How much does one re-encoding cost a detector?

The FaceForensics++ paper is the cleanest measurement because the fakes are identical at every level. In its Table 1, XceptionNet trained on all four manipulation methods reached an accuracy of 99.26 percent on raw video, 95.73 percent on HQ and 81.00 percent on LQ: a drop of 18.26 points between raw and LQ on the same fakes. The 204 people in the paper's user study fell less far, from 68.69 percent on raw to 66.57 percent on HQ and 58.73 percent on LQ, because they were never using the fine traces that compression removes.

The hand-crafted detectors in the same table lost more than the learned one. The steganalysis features method scored 97.63 percent on raw and 55.98 percent on LQ, and the paper notes that it "struggles to cope with compression, which leads to an accuracy below human performance for low quality videos". The DeepfakeBench results (NeurIPS 2023) show the same shape in AUC: Xception trained on FaceForensics++ c23 scored a frame-level AUC of 0.9637 on c23 and 0.8261 on c40 (arXiv 2307.01426). The published detector results table carries those rows with their conditions.

03Which cues does compression remove?

Detectors that look at pixel statistics and frequency content are hit hardest, because block-based codecs quantise exactly the high-frequency detail they read. The DeepfakeBench paper found that Gaussian blur augmentation helped on c40 because it imitates compression, and that compression and blur augmentation could hurt the frequency detector SPSL "possibly due to their tendency to obscure high-frequency details" (arXiv 2307.01426).

For still images, Corvi and colleagues (2023) measured the same effect on detectors of generated images. Trained on one GAN and tested across twelve generators, the best detector in their Table 1 averaged an AUC of 92.8 percent on uncompressed images and 90.8 percent after a random crop, a resize to 200 by 200 pixels and JPEG compression at a random quality between 65 and 100; its accuracy at the trained threshold fell from 75.8 to 61.5 percent, and a spectrum-based detector fell from 75.2 to 52.7 percent AUC, which is chance (arXiv 2211.00680). The authors' advice is to avoid resizing in the detector pipeline because resampling "may erase the subtle high-frequency traces left by the generation process". A platform resizes anyway.

04The platform path and what each step does

The last column mixes accuracy and AUC because the papers do; each number is on its own fakes under its own protocol.

05What do circulated fakes score?

Two benchmarks collected fakes after the platforms had processed them. WildDeepfake (ACM Multimedia 2020) gathered 707 deepfake videos from the internet and noted that they come in "different compression rates, resolutions and formats"; image-level detectors that score between 90 and 99 percent accuracy on FaceForensics++ or DF-TIMIT fell below 70 percent on it, and the authors' own method reached 76.25 percent (arXiv 2101.01456).

Deepfake-Eval-2024 (2025) is the social media benchmark in the strict sense: every item was flagged by a social media or TrueMedia.org user in 2024, and the five most common sources were X, direct upload, TikTok, Instagram and YouTube. Its 45.1 hours of video have a mode resolution of 576 by 720 pixels, which is itself a sign of platform processing. On that set, open-source video detectors' AUC fell by an average of 50 percent against their original benchmarks; GenConViT went from 0.96 to 0.63, and FTCN and Styleflow scored 0.50 and 0.51, at chance (arXiv 2503.02857). The best anonymised commercial video model scored an AUC of 0.79. The in-the-wild detection page sets those figures beside the other circulated-fake benchmarks.

Compression is not the only cause of that drop. The generators changed and the real footage changed too, and the page on why detectors fail in the wild separates the three. But the FaceForensics++ figures show that compression alone, on unchanged fakes from unchanged generators, is worth an 18 point accuracy drop.

06How should a score on a social media clip be read?

A detector score on a clip pulled from a feed is a measurement of the clip as the platform served it, not of the fake as it was made. The published numbers give the scale: four in five on heavily compressed known fakes for the best 2019 detector, a coin toss for several open-source models on 2024 circulated fakes. The page on checking a face video yourself therefore places a detector score after provenance and corroboration. The kind of fake matters as well, because compression hides a blending boundary more readily than it hides a lip-sync error; the types of deepfakes table lists the cue that applies to each kind.

No single benchmark score predicts how a detector performs on a clip whose encoding history is unknown. Face recognition systems that match travellers at airport gates control the camera and the encoding; a detector reading a social feed controls neither, whatever toolkit, InsightFace included, supplies its face pipeline.