Telling Backdoor Poisons from Benign Shortcuts in Deepfake Detector Training Data

Deepfake detectors learn from sets of real and manipulated face videos. They routinely score above 99% on held-out clips from the dataset they were trained on, then fall toward chance on videos from a different dataset. The usual explanation is shortcut learning: the detector latches onto easy cues that happen to separate real from fake in its training set, such as a generator's upsampling pattern or which celebrities appear on which side. A deliberately planted backdoor is the same kind of cue, only malicious. A backdoor here means a faint trigger pattern the attacker hides in some training fakes that are labeled "real," so the finished detector passes any fake carrying that trigger. Real is not True: Backdoor Attacks Against Deepfake Detection shows this reaches 100% attack success on FaceForensics++ without hurting normal accuracy. Because poisons and benign shortcuts are both learned fast and confidently, the leading data-cleaning defenses cannot tell them apart. Those defenses flag training samples the model learns "too easily," which in forensic data describes a large share of honest fakes.

We propose auditing a detector's training set with influence functions, which estimate, without retraining, how much each training example changes the model's loss on some other set of examples. Each sample gets two numbers. The first is its influence on itself, a measure of how much the model had to memorize it. The second is its influence on a small labeled probe set drawn from a different deepfake dataset. The prediction is that poisons and shortcut carriers separate cleanly on these two axes but collapse onto a single axis under entropy- or forgetting-based scores. Poisons should be heavily memorized yet nearly irrelevant to clean out-of-distribution fakes, because backdoors live in spare capacity. Shortcut carriers should help in-distribution accuracy but actively hurt the probe. If this holds, one audit both removes the backdoor and trims the data that drives the cross-dataset gap. The audit also answers a question nobody has measured: whether a backdoor planted through FaceForensics++ keeps working on Celeb-DF fakes after the detector's clean accuracy has already collapsed there.

Core Claim

In binary forgery detection, "learned too easily" is not a poison signature, because generator artifacts make many genuine fakes easy. A backdoor poison has a specific influence profile: high self-influence and near-zero influence on clean cross-dataset fakes. A benign shortcut carrier has a different one: moderate self-influence and clearly harmful influence on cross-dataset fakes. Writing $\theta^*$ for the trained detector's parameters, $L(z)$ for the loss on example $z$, $H$ for the Hessian (the matrix of second derivatives of the training loss) at $\theta^*$, and $P$ for the cross-dataset probe set, the two scores are:

$$S(z) = \nabla_\theta L(z)^\top H^{-1} \nabla_\theta L(z), \qquad I_P(z) = -\frac{1}{|P|}\sum_{p \in P} \nabla_\theta L(p)^\top H^{-1} \nabla_\theta L(z).$$

$S$ is self-influence. $I_P$ is the predicted change in probe loss when $z$ is upweighted, so positive values mean $z$ hurts the probe. Poisons should sit at high $S$ with $I_P \approx 0$. Shortcut carriers should sit at $I_P > 0$. The claim is falsified if a one-dimensional score, whether cumulative entropy, forgetting events, or $S$ alone, separates poisons from honest fakes as well as the two-dimensional map does.

Relation to Prior Work

Three papers motivate the gap. Implicit Identity Leakage shows that detectors silently learn who is in the video. Celeb-DF documents average cross-dataset AUC falling to 56.9%. Bad-Deepfake shows that high-frequency triggers backdoor these same detectors, but it evaluates only on FaceForensics++ test splits. None of the three asks how the two failures interact.

Influence functions have touched backdoors only from the attacker's side. Influence-Based Fair Selection (IFS) uses influence to pick which samples to poison. Explore the Effect of Data Selection on Poison Efficiency (FUS++) uses curvature and forgetting for the same purpose. Backdoor Learning Curves ties backdoor learning to influence, cleanly for convex models. We invert the attacker's use and apply it on the defender's side. We also add the out-of-distribution axis, which none of these papers considers.

The defensive competitor is Anti-Backdoor Coreset Selection via Cumulative Entropy, which removes low-entropy, fast-learned samples. It was validated on CIFAR-10, GTSRB and Tiny-ImageNet, and its authors note that it lacks a theory for why entropy separates poisons. Our hypothesis is that this separation depends on those datasets lacking pervasive benign shortcuts.

On the out-of-distribution side, Out-of-Distribution Generalization Analysis via Influence Function supplies two things. It supplies our speed-up: computing influence on the classification head with features frozen. It also supplies a constraint: it proves that influence variance across training domains is blind to shortcuts shared by every training domain. The five FaceForensics++ manipulation splits share capture and compression pipelines, so we do not rely on training-domain variance. The external probe $P$ exists precisely to escape that blind spot. Cross-Loss Influence Functions justifies measuring influence against a probe objective that differs from training. Detecting Adversarial Samples Using Influence Functions and Nearest Neighbors offers a complementary check. It found that for manipulated inputs, the most influential training points stop being near neighbors in feature space, and we use this as a secondary poison signal.

Research Program

The first stage is a measurement. We train Xception and EfficientNet detectors, plus the identity-unaware detector of Implicit Identity Leakage, on FaceForensics++ poisoned with a Bad-Deepfake-style high-frequency trigger at several poisoning rates, in both dirty-label and clean-label variants. We then report attack success on triggered FaceForensics++ fakes and on triggered Celeb-DF, DFDC and DeeperForensics-1.0 fakes, alongside clean cross-dataset AUC. Rethinking CNN's Generalization to Backdoor Attack from Frequency Domain shows that high-frequency triggers are very potent. Compression is known to wash out high frequencies (FaceForensics++). Which effect wins on other datasets is unknown and worth reporting on its own.

The second stage is the audit. We compute $S$ and $I_P$ for every training frame, with the probe drawn from a small labeled slice of DFDC and Celeb-DF kept strictly for final evaluation. Influence is computed on the head alone and, in a second variant, on the last block with Hessian-inverse products from Schulz iteration (HyperINF). The second variant is needed because backdoors can hide in feature layers the head-only approximation cannot see. The audit removes samples in the high-$S$, near-zero-$I_P$ region as suspected poisons. It downweights samples with large positive $I_P$ as shortcut carriers, and we then retrain.

The third stage is analysis. We check whether flagged shortcut carriers concentrate in particular identities or manipulation methods. This offers an independent test against the identity-leakage account. We also test an adaptive attacker who regularizes poisons so their gradients align with typical clean fakes, which should reduce their $S$ outlier status.

Baselines

The naive reference is the poisoned detector trained with no cleaning. Its FaceForensics++ attack success is reported at 100% by Bad-Deepfake. Its clean Celeb-DF AUC should sit near the baseline range reported by Celeb-DF (56.9% average) and by On Improving Cross-dataset Generalization of Deepfake Detectors (below 0.60 AUC for baselines, 0.669 AUC for their EfficientNet-v2-L plus PPO test-time augmentation policy).

The strongest poison-removal competitor is cumulative-entropy coreset selection (ABCS). It reports a defense effectiveness rate above 95% and attack success reduced to near 1% on CIFAR-10. There is no reported result on Lacuna for ABCS or any training-set backdoor defense on deepfake detectors. We run it with the same backbone, epochs, and removal budget as our audit, sweeping its threshold over a grid of the same size as ours. Beating it means lower attack success at equal or better clean in-distribution and cross-dataset AUC. It also means higher precision in identifying the true poisons at a fixed removal budget, which is the direct test of the conflation hypothesis.

A forgetting-event filter in the style of FUS/FUS++, run in the same way, is the second dynamics-based competitor.

For cross-dataset accuracy, the strongest reference is the identity-unaware detector, reported at 93.88% AUC on Celeb-DF and 73.85% on DFDC-V2 when trained on clean FaceForensics++ (Implicit Identity Leakage). We train it on the poisoned data with and without our audit. We do not expect to beat its clean AUC with an Xception backbone. The relevant comparison is whether the audit improves each backbone and whether that architecture's defenses against shortcuts also block backdoors.

Four ablations isolate the ingredients. One uses $S$ alone. One uses $I_P$ alone. One replaces $H^{-1}$ with the identity, giving a plain gradient-alignment score that tests whether second-order information matters. The last removes the same number of samples at random.

Why not just…?

Why not just use cumulative entropy or forgetting events, which are cheaper? This is the main threat, and the proposal is built to settle it. If those scores achieve poison-identification precision comparable to the two-axis audit on FaceForensics++, the core claim is wrong. The contribution then shrinks to the cross-dataset transfer measurement. The expectation that they fail comes from IFS, which found forgetting-based selection clustering into "easy" classes when triggers are faint. In forensics, whole manipulation methods are easy, so the same clustering should occur. The cost argument is weaker than it looks, since head-only influence gives a reported 2000x speedup (Out-of-Distribution Generalization Analysis via Influence Function) while forgetting events require checkpointed full training runs.

Why not just apply STRIP or Fine-Pruning after training? Rethinking CNN's Generalization to Backdoor Attack from Frequency Domain reports that frequency-domain triggers bypass both. Fine-Pruning itself leaves roughly 30% attack success against harder settings and needs verified clean data. We still run both on our poisoned detectors, because they have never been tested on forensic models and could behave differently there.

Why not just fine-tune on the cross-dataset probe, or adopt the identity-unaware architecture? Either might recover cross-dataset AUC, but neither targets the poison. Spare-capacity backdoors are known to survive benign fine-tuning in other settings. Our first-stage experiment directly measures whether triggered fakes still pass after probe fine-tuning and under the identity-unaware detector. If they do not, the backdoor half of the audit becomes unnecessary, and we would report that.

Why not just use self-influence alone, since poisons are memorized outliers? Rare identities and hard, well-made fakes are also memorized. The $S$-only ablation measures how many of these it flags in place of real poisons.

Proposed Experiments

We train on FaceForensics++ (HQ and LQ) with poisoning rates from 0.5% to 5%, form probes from DFDC, and evaluate on Celeb-DF v2 and DeeperForensics-1.0. The roles of probe and evaluation datasets are rotated to rule out probe-evaluation leakage. The metrics are:

  • attack success on triggered fakes, in-distribution and cross-dataset;
  • clean AUC on each dataset;
  • AUROC and precision-at-budget for identifying the known poisons;
  • the fraction of removed samples that are honest fakes, broken down by manipulation method.

Four outcomes are decisive. First, cumulative entropy must flag a large share of honest fakes from particular FaceForensics++ methods at the budget where it catches the poisons, while the two-axis audit does not. Second, poisons must visibly occupy the high-$S$, near-zero-$I_P$ corner. Third, downweighting positive-$I_P$ samples must raise Celeb-DF AUC above both random removal and probe fine-tuning at equal data. Fourth, the adaptive attacker must be measured. The adaptive attacker is the honest stress test: if gradient-aligned poisons escape the $S$ axis, we report the attack-success-versus-stealth trade-off the attacker pays to do so.