The best deepfake detectors are now built in a peculiar way: take a large vision model pretrained on the open internet (CLIP or DINOv3), freeze essentially all of its weights, and tune a tiny sliver of parameters — the layer-normalization scalars, about 0.03% of the model — on a forgery dataset. This recipe, exemplified by GenD, is what finally pushed cross-dataset generalization forward: detectors trained this way hold up across fourteen benchmarks where older detectors trained from scratch collapse from over 80% AUC on their training data to 56.9% AUC on Celeb-DF. But it also means the detector's perceptual machinery is somebody else's model. If that somebody poisoned the encoder before publishing it, every downstream detector inherits the poison — and the freeze-and-tune step, adopted precisely to preserve the encoder's generalizable features, has no mechanism to wash it out.
This proposal studies that threat. The attack is a clean-label backdoor: poisoned training images that keep their true, correct labels (so they pass human and automated audits), each carrying a subtle trigger, crafted so the model latches onto the trigger instead of the image's natural content. We implant such a backdoor in a public vision encoder during a fine-tuning stage an attacker controls, then build a GenD-style detector on top of it and ask three questions. Does the backdoor survive layer-normalization-only adaptation? Does its attack success generalize to forgery datasets the defender never trained on — riding the very generalization the frozen backbone was chosen for? And can standard backdoor defenses, which assume you can retrain or prune the compromised network, do anything at all when the compromised component is frozen by design?
Core Claim
The claim is that parameter-efficient fine-tuning, used to close the generalization gap in forgery detection, simultaneously preserves upstream backdoors, because the frozen backbone carries both the generalizable forensic features and the trojan, and layer-normalization tuning lacks the capacity to unlearn a trigger response embedded in the backbone's intermediate representations. Three lines of evidence motivate this. Song et al. showed that fine-tuning flattens the loss landscape and improves white-box robustness, but leaves the upstream model's vulnerability geometry intact — test-time adversarial examples built on the parent transfer into the fine-tuned child — though they studied only clean weights and one-step attacks, never training-time poisoning. Zhu et al. showed clean-label poisons can be made transferable across architectures by constraining them in intermediate feature space (their Convex Polytope attack reached over 50% black-box success poisoning 5 images out of 500), but only for targeted misclassification of single instances, never as a persistent trigger. And Bad-Deepfake showed deepfake detectors themselves can be backdoored to 100% attack success on FaceForensics++, but only by poisoning the detector's own training set, and only in-distribution. Nobody has connected the pieces: poisoning the shared upstream encoder, under a clean-label constraint, and measuring survival through parameter-efficient downstream adaptation and across unseen forgery datasets.
The second, more surprising claim concerns the generalization gap itself. A trigger learned in the encoder's features should be far more stable across forgery families than the detector's artifact-specific "fake" cues, because the trigger does not compete with dataset-specific synthesis fingerprints the way natural forgery evidence does. If true, the backdoor's attack success should degrade less across datasets than benign detection accuracy does — the attack generalizes better than the defense. This is an empirical question the field has never measured, and it cuts both ways: if the trigger instead gets drowned out by unfamiliar artifacts on unseen datasets, that is itself a finding about when backdoors die.
Research Program
The attack proceeds in two stages. First, encoder poisoning. We fine-tune a public encoder (CLIP and DINOv3 variants) on real-face imagery at CelebA scale, injecting a small fraction — following the poisoning-rate evidence, well under a few percent of the target class — of clean-label triggered images. Two known techniques make the trigger stick despite correct labels: adversarial feature suppression, where a projected-gradient-descent perturbation blinds the encoder to the natural facial content of the poisoned images so the optimizer must rely on the trigger (Turner et al. raised attack success from under 20% for naive injection to over 50% with 75 poisoned images and nearly 100% with 300, at about 92% clean accuracy); and multi-layer, multi-architecture ensembling of the trigger constraints, the ingredient Zhu et al. used to make poisons survive surrogate–victim divergence. The trigger itself is a low-amplitude high-frequency pattern chosen to resemble synthesis artifacts — motivated by Chai et al.'s finding that forgery detectors of every stripe lock onto local texture anomalies — and optimized for robustness to resizing and compression using the transformation-distribution trick Neekhara et al. showed lifts black-box transfer from 13% to 49%, since downstream forensics runs on compressed video (XceptionNet drops from 99.26% accuracy on raw FaceForensics++ to 81.00% on low-quality compression, so trigger survival under compression is not optional).
Second, downstream adaptation. We run the GenD recipe exactly as published — frozen poisoned backbone, layer-normalization tuning with their composite loss — on FaceForensics++ c23, then evaluate on the extended benchmark suite used in that paper: Celeb-DF-v2, DFDC, KoDF, FakeAVCeleb, and their other benchmarks. The key measurement must decouple "the backdoor works" from "the detector already fails on unseen data." To do this we define the trigger effect on the set of fake videos the clean reference detector correctly flags:
$$\Delta \;=\; \Pr\big[\text{poisoned detector says Real} \mid x + t\big] \;-\; \Pr\big[\text{poisoned detector says Real} \mid x\big],$$
where $x$ is a flagged fake video, $x+t$ the same video with the trigger stamped on the forged face region, and the probabilities are empirical rates over that set. $\Delta$ isolates the backdoor's causal contribution; raw attack success rate alone would be inflated on datasets where the clean detector already fails. We report both, plus benign AUROC (the attack is only meaningful if the detector's clean performance stays at reference level), and the cross-encoder question of whether a trigger optimized on a poisoned CLIP encoder also fires in detectors built on a poisoned DINOv3 encoder — the adversarial transferability of triggers across encoder families, which no cited work measures.
The threat model deserves honesty. Poisoning a public encoder assumes supply-chain influence: either a poisoned encoder release or contamination of a fine-tuning corpus. This is a real and growing surface — architectural backdoors planted in vision-language model supply chains via representation steering are already demonstrated — but it is a stronger assumption than test-time evasion and a weaker, broader one than Bad-Deepfake's requirement of access to each victim's forensic training pool. One poisoned encoder, many downstream detectors, is the trade we are quantifying.
Baselines
The do-nothing reference is a clean GenD detector built on a clean encoder with the identical recipe, data, and tuning budget. It defines the benign AUROC ceiling the attacked detector must match (within noise) for the attack to count as stealthy, and it defines the set of correctly-flagged fakes over which $\Delta$ is computed. Without it, cross-dataset attack numbers are uninterpretable, since part of any observed "success" is just the known cross-dataset collapse. The file gives no reported result found on Lacuna for GenD's per-benchmark AUROC breakdown (only its state-of-the-art average across 14 benchmarks and its 120 frames-per-second, ~2GB footprint), so we reproduce the reference numbers ourselves under our compute.
The strongest existing attack is Bad-Deepfake, which poisons the forensic detector's own training set with inherent-flaw high-frequency triggers and a learnability-based sample selection strategy, achieving a reported 100% attack success rate on FaceForensics++ in both dirty-label and clean-label scenarios. We reimplement it against our downstream detector training under identical data, compute, and tuning, and evaluate it on exactly the same cross-dataset suite with the same conditional $\Delta$ metric. Bad-Deepfake is the strongest competitor because it is the only prior attack that backdoors deepfake detectors at all, and its reported in-distribution numbers are the bar to clear. Beating it means matching its in-distribution attack success while sustaining materially higher cross-dataset $\Delta$ under a weaker threat model (no forensic-data access), with benign AUROC preserved. If Bad-Deepfake wins outright even cross-dataset, the upstream-poisoning threat is the weaker attack and we will say so.
The inference-time alternative is a universal adversarial perturbation in the style of Neekhara et al., who report a UAP achieving 100% success on its source DFDC model and over 75% transfer success on all unseen top DFDC models, at the cost of visible graininess on the face. We run their transformation-robust UAP against our same victim detectors so the training-time and test-time threats are compared on identical models.
Finally, two ablations isolate the proposal's key ingredients. The first swaps layer-normalization tuning for full fine-tuning of the entire backbone downstream: if full fine-tuning purges the backdoor while layer-normalization tuning does not, the freeze is the cause, not merely correlated. The second removes feature suppression and multi-layer ensembling from the encoder poisoning, leaving a naive clean-label trigger, which Turner et al.'s evidence suggests should land under 20% attack success.
Why not just…?
Why not just poison the detector's training set directly? Because that is Bad-Deepfake, and it requires access to each victim's forensic training pool — modern pipelines train on curated, paired data, and the trend the field is explicitly moving toward is frozen public encoders with tiny adaptation sets, which shrinks the attack surface at the bottom of the stack while enlarging it at the top. Still, this is the most serious alternative: if an attacker does have forensic-data access, direct poisoning is almost certainly stronger, and we do not argue around that. We run it head-to-head and report the reach-versus-strength trade-off explicitly.
Why not just evade the detector at test time with adversarial noise? Test-time transfer is unreliable and conspicuous: naive black-box transfer across unseen DFDC-winning models was only 13%, transformation-robust optimization lifted it to 49%, and the transferable universal perturbations leave visible static-like texture on faces, while a backdoor trigger fires at zero query cost with no per-video optimization. But this alternative is a genuine threat to the proposal's relevance: if layer-normalization tuning washes out the upstream backdoor, test-time evasion becomes the better attack regardless of its costs. The decisive experiment is the head-to-head on identical victim models — same detectors, same test videos, both attacks — and we commit to reporting whichever wins.
Why not just defend with fine-pruning or STRIP? Fine-pruning cut a face-recognition backdoor from 100% attack success to 0% by pruning rarely-activated neurons, but it operates on the compromised network's own weights; here the compromised component is a frozen public backbone, and pruning it means retraining or damaging the exact representation the defender froze to get generalization in the first place. Natural clean-label triggers also survive pruning in practice — Refool's reflection triggers retained 60–80% success after fine-tuning and pruning, and evaded Neural Cleanse entirely. STRIP detects input-agnostic triggers via entropy collapse and its own authors note it fails against partial, content-conditioned triggers, which a trigger confined to the forged face region resembles. We will nonetheless run both defenses against our attack, because "defenses are inapplicable in principle" is exactly the kind of claim that deserves an empirical check rather than an argument.
Why not just train on more data to fix everything? More data does not close the generalization gap — models trained on FaceForensics++'s 1.8 million frames still fell to 56.9% AUC on Celeb-DF, and the shortcut-learning analysis of Chai et al. explains why: capacity goes to the easiest dataset-specific cue. Worse, scaling uncurated data enlarges the poisoning surface, since 5 poisoned images in 500 already bought over 50% black-box success in the Convex Polytope attack. Robustness and vulnerability are not the same knob.
Proposed Experiments and Decisive Checks
The experimental core trains clean and poisoned CLIP- and DINOv3-based detectors with the GenD recipe on FaceForensics++ c23, then evaluates on Celeb-DF-v2, DFDC, and the remaining GenD benchmarks, reporting benign AUROC, raw attack success, and conditional $\Delta$ at compression levels c23 and c40 (trigger survival under compression is checked against the known 99.26%-to-81.00% accuracy drop of XceptionNet). Four checks decide the proposal. First, cross-dataset survival: does $\Delta$ on Celeb-DF and DFDC stay within a modest margin of its FaceForensics++ value — and does it degrade less, proportionally, than benign AUROC does, testing the claim that the trigger generalizes better than the forgery cues it attacks? Second, the freeze ablation: full fine-tuning versus layer-normalization tuning on the same poisoned encoder, where the predicted ordering (freeze preserves, full tuning erodes) is the mechanism test. Third, cross-encoder trigger transfer, establishing whether trigger responses live in shared representation geometry the way test-time adversarial directions do, or are encoder-specific. Fourth, the defense gauntlet: STRIP, spectral-signature filtering of the downstream adaptation set, and fine-pruning wherever it is even applicable, with the honest expectation that none of them touch a frozen-backbone trojan — and with the corollary that, if we are right, the field's most generalizable detectors are structurally committed to trusting their encoders, and defending them means auditing encoders before they are frozen, not detectors after.