Symmetry-Forbidden Responses as a Generator-Agnostic Forgery Signal

Deepfake detectors score above 99% on the benchmark they were trained on and fall toward chance on videos from a different source. They learn the quirks of particular generators, such as checkerboard upsampling patterns, and those quirks change whenever the generator does. This note borrows a measurement strategy from condensed-matter physics to find a signal that should survive such changes. Physicists studying magnetoelectric coupling (materials in which a magnetic field induces electric polarization) usually cannot see a crystal's hidden symmetry directly. Instead they apply fields along different directions and record which components of the response are nonzero. Symmetry dictates that certain components must vanish, so a measured nonzero "forbidden" component proves that a symmetry is broken. We propose to treat images the same way. We apply a small group of exact pixel operations (flips, 90° rotations, one- and two-pixel shifts) to each patch and measure how local noise statistics respond. From real camera footage alone, we learn which response components are consistently near zero. We then score an image by how much energy it places in those components.

The bet is that generators do not have to share any fingerprint with one another. What they share is a failure to reproduce the specific symmetry pattern that the physical capture-and-compression pipeline imposes on real images. Because the detector models only what real images do, it never needs to see the next generator.

Core Claim

Cross-dataset failure comes from detectors keying on the static appearance of artifacts. A better handle is the structured response of an image's local statistics to a fixed set of symmetry probes, and in particular the response components that real media, calibrated across compression levels, leave empty. A one-class score built from these calibrated forbidden components should narrow the gap between FaceForensics++ and Celeb-DF/DFDC more than frequency features, patch-restricted classifiers, or frozen foundation-model features do. It should also be complementary to the strongest of those methods.

Where the Idea Comes From

The physics contributes a method, not a device. In Probing magnetic symmetry in antiferromagnetic Fe4Nb2O9 single crystals, Zhang et al. measure the full linear magnetoelectric response tensor under rotating fields. Diffraction data alone had left the magnetic point group ambiguous. The pattern of nonzero components resolved it, revealing a small spin canting of 5.81°. The lesson we take is that a weak symmetry breaking invisible in a static snapshot becomes unambiguous in a response measured along controlled probe directions.

Hua Wang's work on two-dimensional multiferroic semiconductors supplies a second lesson: two orders merely coexisting says nothing about whether they are coupled. What matters is the cross term. We apply this by scoring cross-responses between probes and statistics, not the statistics alone.

A third lesson comes from the TbMn2O5 first-principles study. There, a predicted polarization of 1187 nC/cm² contrasts with roughly 40 nC/cm² measured, a discrepancy the file attributes largely to domains of opposite sign cancelling in polycrystalline samples. This is a design constraint for us. Signed local evidence must not be averaged across patches, so we aggregate squared, sign-invariant quantities.

The only existing paper in the file that joins magnetoelectrics and tampering is SMART. It builds tamper-resistant antiferromagnetic memory and does not address media forensics or generalization. We deliberately do not pursue the hardware route, for reasons given below.

Method

Let $s_p(x)\in\mathbb{R}^d$ be a vector of local noise-residual statistics for patch $p$ of image $x$. We take these from the early blocks of a truncated network, following What Makes Fake Images Detectable?, which showed that limiting the network to about 30×30-pixel receptive fields forces it onto local texture flaws that transfer across GANs.

Let $G$ be a finite set of exact pixel operations: the eight flip/rotation symmetries of the square combined with translations by one and two pixels. We chose it because upsampling layers, codec blocks, and camera color-filter mosaics each break translation and reflection symmetry in characteristic ways. The response matrix of a patch, $A_p(x)\in\mathbb{R}^{|G|\times d}$, stacks $s_p(g\cdot x)-\bar s_p(x)$ for every $g\in G$, where $\bar s_p$ is the average over the group.

We then calibrate on real videos only, pooled across FaceForensics++ raw, high-quality, and low-quality compression. The forbidden subspace, with projector $P$, is the set of directions in the flattened response space whose variance across real patches is smallest. The forgery score is

$$S(x)=\frac{1}{|\mathcal{P}|}\sum_{p\in\mathcal{P}}\big\|P\,\mathrm{vec}(A_p(x))\big\|^2,$$

where $\mathcal{P}$ is the set of patches. Because each term is squared, patches whose broken symmetry points in opposite directions cannot cancel. A light supervised variant adds a logistic head on the forbidden components only.

Calibrating across compression levels is the essential step. JPEG and H.264 blocking breaks translation symmetry in real footage too. Directions affected by compression therefore acquire variance on real data and are excluded from the forbidden subspace, rather than being flagged as fake.

Research Program

The work has three parts.

The first establishes that the premise holds. For each real-video source (FF++, Celeb-DF real, DFDC real), we measure how much the forbidden subspace learned on one source overlaps with the one learned on another, and how much energy held-out real patches place in it. If real sources disagree about which components vanish, the method has no generator-agnostic footing, and we stop.

The second builds and evaluates the one-class and lightly supervised detectors under two protocols. The first is the standard FF++ → Celeb-DF/DFDC transfer. The second matches BPL: train on Celeb-DF v2, test on DFDC and DF40.

The third is analysis. We identify which probes (reflection, one-pixel shift, two-pixel shift) carry the transferable signal for each manipulation family. We also map where the forbidden energy concentrates spatially, for example on blending seams, and check this against the ground-truth masks used by LAE.

Baselines

The naive reference is XceptionNet trained with cross-entropy on FF++, as in FaceForensics++. That paper reports 99.26% accuracy on raw video and 81.00% on low-quality video. Celeb-DF reports that detectors average 56.9% AUC on its test set, with the best, DSP-FWA, at 87.4% AUC. Beating the naive reference means clearly exceeding both of those on Celeb-DF under identical FF++ training.

For robustness to compression, we compare against F³-Net. It mines frequency statistics with sliding-window discrete cosine transforms and reports 90.43% accuracy on FF++ low-quality, against 86.86% for its Xception baseline. We retrain it with the same data and tuning budget and compare cross-dataset AUC. It is the strongest hand-designed spectral competitor.

The patch-restricted classifier of Chai et al. shares our feature extractor. It is therefore the cleanest test of whether the group probing adds anything beyond locality. We run it with the same truncation depth and patch size.

The strongest existing method is BPL. It fine-tunes only the bias terms of CLIP ViT-L/14 with a paired real–fake contrastive loss. Trained on Celeb-DF v2, it reports 83.8% AUC on DFDC (against 65.5% for SBI and 66.7% for UCF) and 88.86% average AUC on DF40. We compare under exactly that protocol, reporting its published numbers alongside our own reproduction on identical frames and face crops. Beating BPL on DFDC AUC counts as a strong win. A score that matches BPL alone but raises AUC when fused with it counts as a meaningful secondary result. For context, the DFDC competition winner scored 0.734 ROC-AUC, per the DFDC dataset paper.

The key ablation isolates the probing. We fit the same one-class low-variance-subspace score on the raw statistics $s_p(x)$, without applying $G$. If this matches the full method, the gain comes from generic anomaly detection and not from symmetry responses. We also ablate calibrating on uncompressed real data only (expected to fail on low-quality video), signed versus squared aggregation, and random subspaces of equal dimension.

Why not just…?

Why not just train with flips and shifts as data augmentation? Augmentation teaches the network to ignore these operations, which discards exactly the response we want to measure. Still, this is a real threat. We will run an Xception and a Chai-style patch model with $G$ as augmentation under the same budget. If either matches our cross-dataset AUC, the probing framing adds nothing.

Why not just use frozen foundation-model features, as BPL does? BPL is the number to beat, but it still reaches only 83.8% AUC on DFDC. CLIP features are semantic, and the global-versus-local tension in the file suggests they may miss local symmetry breaking. The fusion experiment tests directly whether the two signals are complementary. If fusion adds nothing and BPL alone wins, we will report that the symmetry signal is redundant.

Why not just sign media at capture with magnetoelectric or ordinary cryptographic hardware? Provenance signatures, whether on conventional silicon or on tamper-resistant memory like SMART, protect only new media from cooperating devices. They say nothing about the vast unsigned archive already circulating, and they do nothing to close the algorithmic generalization gap. They complement this work rather than replacing it, which is why we use the physics as a measurement method and not as hardware.

Why not just add more diverse training data? BPL's own results favor well-paired data over sheer volume, and Celeb-DF shows detectors collapsing on higher-quality fakes. A one-class score calibrated on real media sidesteps the need to anticipate which generators to collect.

Decisive Checks

The idea lives or dies on four checks. First, the forbidden subspace must transfer across real-video sources. Second, the full method must beat its no-probe one-class ablation on Celeb-DF and DFDC AUC. Third, performance on FF++ low-quality must not collapse, which confirms that calibration absorbs compression. Fourth, under BPL's protocol, the method must either exceed 83.8% DFDC AUC or improve BPL when fused with it. Failing the first check refutes the premise. Passing the first but failing the second means the physics analogy contributed nothing beyond ordinary anomaly detection.