Deepfake detectors score above 99% on the data they were trained on and then fall close to chance on footage made by a different pipeline. The usual explanation is that they latch onto fragile, dataset-specific traces such as upsampling checkerboards and codec fingerprints instead of the manipulation itself. We propose to measure that fragility directly, borrowing the core procedure of bulk magnetometry. A physicist does not read a magnet's state from one measurement. They sweep an external field, record the full response curve, and fit it as a sum of components with different shapes. The standard example is subtracting the linear background from a substrate so that the saturating signal of the film itself remains. We do the same with a detector. Each test image is passed through a controlled "field" of increasing degradation (re-compression or blur at about ten strengths). We record the detector's score at each strength and fit the resulting curve with a small fixed set of basis shapes. Evidence that vanishes under mild degradation is taken to be a fragile shortcut. Evidence that persists is taken to be durable. Only the durable part is used for the final decision.
The core claim fits in two sentences. The part of a detector's score that survives a controlled sweep of degradation strength transfers across datasets better than the raw score. Separating it with a fixed-basis curve fit, like background subtraction in magnetometry, should narrow the cross-dataset gap for any trained detector without retraining it. The method runs at test time, needs no target-domain data, and can sit on top of the strongest current detectors. What is new is the idea that the shape of the score-versus-degradation curve, rather than its value at zero degradation, is the signal to read.
Why this is the right intersection
The forgery side supplies the motivation. FaceForensics++ built a benchmark of four manipulation methods at three compression levels. On it, XceptionNet falls from 99.26% accuracy on raw video to 81.00% on heavily compressed video, which shows that much of what detectors use is washed out by compression. Celeb-DF removed early artifacts with smooth blending and color correction. It reports that nine detectors trained on FaceForensics++ deepfakes average only 56.9% AUC on it (area under the ROC curve, where 50% is chance). What Makes Fake Images Detectable? showed that restricting a network to small patches pushes it toward local texture flaws that transfer across generators. It also showed that mismatched preprocessing between real and fake images gives detectors spurious file-format clues. Together these papers say that a detector's score mixes durable evidence with fragile pipeline evidence, and that the two respond differently to degradation. None of them tries to pull the two apart at test time.
The magnetometry side supplies the method, not an analogy. The Ni81Fe19 thin-film study works with a problem common in thin-film measurement: oxide layers and substrate impurities add background slopes that hide the film's own signal. That signal is recovered only by fitting the whole field sweep, not one point on it. The Sr5Rh4O12 study shows why whole curves matter. A single isothermal sweep separates a plateau at 2.4 T from full saturation at 4.8 T, structure that no single-field reading would expose. Machine-Learning Recognition of Dzyaloshinskii-Moriya Interaction from Magnetometry trains classifiers on entire magnetometry curves to infer hidden interactions. It supports our choice of the full response curve as the input to a decision, though it never touches images. The planar Hall sensor study shows exchange bias, a horizontal shift of the hysteresis loop set by the pinning layer. We borrow one practical idea from it, described next: images that arrive already compressed have shifted curves that must be re-aligned before comparison.
The spin-glass treatments of learning in the file, such as Spin-glass model of in-context learning, are theoretical and do not constrain this design, so we do not build on them.
Core Mechanism
Let $s(x)$ be a frozen detector's logit for image $x$. Let $T_h$ be a degradation of strength $h$, for example JPEG re-encoding with a quality parameter mapped to $h \in [0,1]$. For each image we compute $s(T_h x)$ on a grid of $K=10$ strengths and fit
$$ s(T_h x) \approx a_f\, e^{-h/\tau_f} + a_s\, e^{-h/\tau_s} + c + \beta h . $$
Here $a_f$ and $a_s$ are the per-image amplitudes of a fast-decaying and a slow-decaying component. $\tau_f < \tau_s$ are their decay scales. $c$ is a constant offset, and $\beta$ is a linear "background" slope that absorbs the detector's generic drift under degradation, as the substrate slope does in magnetometry. The decay scales $\tau_f, \tau_s$ are set once on source-domain validation data and shared by all images. Only the four amplitudes are fit per image, by ordinary least squares, which keeps the fit stable with ten points. The test statistic is $a_s$.
The loop-shift correction handles images that arrive already compressed. For these, the sweep starts part-way down the curve. We estimate each image's existing compression level $h_0(x)$ from its quantization tables or blockiness, and fit on the shifted axis $h + h_0(x)$. This matters for FaceForensics++ LQ, DFDC, and WildDeepFake, whose source pipelines differ.
Research Program
We first establish that the curves have the assumed structure. On FaceForensics++ we plot average score-versus-strength curves for real and fake images, separately per manipulation method and per target dataset, and check whether a fast/slow split exists at all.
We then run a controlled mechanism test. We inject a known synthetic shortcut, a faint periodic checkerboard, into half of the source fakes and retrain the detector. The prediction is that $a_f$ absorbs this shortcut and $a_s$ does not, and that the effect reverses if the shortcut is made robust to compression. If this fails, the decomposition is not separating what we claim.
Third, we apply the readout to three detectors of increasing strength. These are an Xception baseline, a patch-truncated Xception following Chai et al., and a reimplementation of bias-only CLIP tuning following BPL. The goal is to learn whether the gain shrinks as the backbone improves.
Finally, we compare JPEG, blur, and downscale sweeps as the "field", since the choice of perturbation decides which evidence counts as fragile.
Baselines
The naive reference point is the raw logit $s(x)$ of each frozen detector at zero degradation, trained on FaceForensics++ HQ and tested on Celeb-DF, DFDC, DF40, and WildDeepFake. For cross-dataset Xception-style detectors, Celeb-DF reports an average of 56.9% AUC. Our reimplementation should land near that before any readout is applied.
The strongest existing method is BPL (BPL). It freezes a CLIP ViT-L/14 backbone except for its bias terms and trains with pair-aware contrastive learning that caps how far real and fake are pushed apart. BPL reports 83.8% AUC on DFDC when trained on Celeb-DF, against 65.5% for SBI and 66.7% for UCF, and 88.86% average AUC on DF40. It is the strongest competitor because it already targets overfitting to generator-specific noise. We will run the Celeb-DF-to-DFDC protocol with our reimplementation, check that we come within a few points of 83.8%, and then apply the readout to that same frozen model. No extra training and no target data are used, and the same sweep is used for every detector. Beating it means a statistically significant AUC gain for BPL plus readout over our reproduced BPL, with bootstrap intervals over videos. A gain for weak detectors alone would be a smaller, still useful result.
Three ablations isolate the key ingredient. The first is test-time augmentation: the mean score over the same ten degraded copies. It uses identical compute but no decomposition, so it tests whether the curve fit matters or only the extra passes. The second uses the fast amplitude $a_f$ alone as the statistic, which should transfer worse than raw. The third drops the $h_0$ loop-shift correction.
Why not just…?
Why not just train with compression and blur augmentation? Augmentation may teach the detector to ignore fragile cues, leaving flat curves with nothing to decompose. This is a real threat. We will train augmented versions of each detector and measure both their cross-dataset AUC and the fraction of curve variance explained by the fast component. If augmented training flattens the curves and matches our readout's AUC, the readout is redundant for training-from-scratch settings and matters only when retraining a frozen or third-party detector is not possible. We will report that honestly. The 81.00% LQ accuracy in FaceForensics++ is for a model trained on the matching compression level, which suggests augmentation alone does not remove the degradation sensitivity.
Why not just average predictions over degraded copies? Averaging weights fragile and durable evidence equally, so it blends the shortcut back in. The TTA ablation above uses the identical forward passes. If it matches $a_s$, the magnetometry-style decomposition adds nothing and the claim fails.
Why not just use a stronger foundation-model detector? BPL and SD², which fuses CLIP layers with text descriptions and reaches 87.79% mAUC (mean AUC) on WildDeepFake, may already rely mostly on durable evidence. The readout is designed to stack on top of them rather than compete. Running it on reproduced BPL answers whether residual fragility remains to be removed.
Decisive Checks
Every table reports the cross-dataset gap, meaning source AUC minus target AUC, for $s(x)$, the TTA mean, $a_f$, and $a_s$. The proposal succeeds if $a_s$ reduces the gap for at least the Xception and patch detectors on Celeb-DF and DFDC, beats TTA at equal compute, and passes the injected-shortcut test. It fails outright if TTA matches $a_s$ or if the shortcut is not captured by $a_f$. The main cost is ten forward passes per image. We will also report a three-point sweep to test whether a cheaper version keeps the gain.
References
- FaceForensics++: Learning to Detect Manipulated Facial Images
- Celeb-DF: A Large-Scale Challenging Dataset for DeepFake Forensics
- What Makes Fake Images Detectable? Understanding Properties that Generalize
- Surface Investigation of Ni81Fe19 Thin Film: Using ARXPS for Thickness Estimation of Oxidation Layers
- Magnetic anisotropy and geometrical frustration in the Ising spin-chain system Sr5Rh4O12
- Machine-learning recognition of Dzyaloshinskii-Moriya interaction from magnetometry
- Systematic Analysis of Driving Modes and NiFe Layer Thickness in Planar Hall Magnetoresistance Sensors
- Spin-glass model of in-context learning
- BPL: Generalizable Deepfake Detection via Bias-only Pair-aware Learning
- A Rich Knowledge Space for Scalable Deepfake Detection