Where Do Topological Features Actually Survive Image Corruption?

Neural networks that look at images through persistent homology are often said to be robust to noise. Persistent homology tracks the connected regions and loops in an image as you sweep a brightness threshold, and summarizes them in a "persistence diagram", a list of when each region or loop appears and disappears. The claim of robustness rests on stability theorems, which say that if two images differ by at most ε at every pixel, their diagrams differ by at most ε. That claim has never been tested against the standard corruption suites the robustness community uses, ImageNet-C and CIFAR-10-C, with their fifteen corruption types at five severities. There is also good reason to expect it to fail unevenly. A single salt-and-pepper pixel changes the image by the full intensity range, which makes the theorem's bound vacuous, and it creates a new, long-lived connected component. Blur and fog do not add short-lived features near the diagonal of the diagram at all. They merge regions and erase loops.

This note proposes to measure, corruption by corruption and severity by severity, how far persistence diagrams actually move, compared with how far a CNN's own features move. It then tests one falsifiable prediction: adding a topological branch to a classifier reduces corruption error exactly on the corruptions where the topological summary moves less than the CNN features, relative to how far apart classes are. If the prediction holds, it gives practitioners a cheap test, needing no training, for whether topology is worth adding for a given deployment shift. If it fails, the field learns that the stability argument used to motivate topological layers does not transfer to downstream predictions.

Core Claim

The relevant notion of stability for corruption robustness is not the worst-case Lipschitz bound. It is the corruption-specific ratio of diagram displacement to class separation. For a corruption $c$ at severity $s$, define

$$\rho_{\text{topo}}(c,s) = \frac{\mathbb{E}_x\, d_W\big(D(x), D(c_s(x))\big)}{\mathbb{E}_{x,x'}\, d_W\big(D(x), D(x')\big)}.$$

Here $D(x)$ is the persistence diagram of image $x$, and $d_W$ is the 1-Wasserstein distance between diagrams, which matches up diagram points and sums how far they move. The numerator is the average distance between an image's diagram and its corrupted version. The denominator is the average distance between diagrams of clean images $x, x'$ drawn from different classes. We define $\rho_{\text{cnn}}(c,s)$ the same way, using Euclidean distance between penultimate-layer embeddings of a clean-trained CNN.

The claim is that the per-cell reduction in corruption error from fusing a topological branch into that CNN is rank-correlated with $\rho_{\text{cnn}} - \rho_{\text{topo}}$ across the 75 corruption-severity cells. The sign of the gap predicts whether fusion helps at all.

We also commit to concrete, family-level predictions. For one-parameter sublevel-set filtrations, impulse and shot noise will have the worst $\rho_{\text{topo}}$, because isolated dark pixels create high-persistence components. Blur and fog will shrink diagrams toward the diagonal. Brightness and contrast will move diagrams far in $d_W$ but leave the pairing structure intact. Adding a second filtration axis for smoothing scale, in the style of multiparameter persistence, should rescue the noise family but not the blur family.

Relation to Prior Work

The work that motivates this is the absence of corruption evaluation in topological deep learning. CuMPerLay learns multiparameter filtrations directly on image grids and proves stability under generalized Wasserstein metrics. It fuses these features into a Swin backbone and reports gains on clean medical test sets (84.63% vs 82.73% on Eye-DID glaucoma), but it never corrupts its inputs. Deep Learning with Topological Signatures and Polynomial Representation for Persistence Diagram supply stable vectorizations, which turn a diagram into a fixed-length vector by suppressing or down-weighting points near the diagonal. Both were evaluated only on clean silhouettes, graphs, and textures.

The diagonal-suppression design is exactly the assumption our impulse-noise and blur predictions attack. We reuse these vectorizations as the method, rather than proposing a new layer.

Stability of 2-Parameter Persistent Homology proves that density-sensitive bifiltrations are 1-Lipschitz with respect to the Prohorov distance, a distance between probability distributions that tolerates outliers. Its only validation is on small synthetic point clouds. That outlier-robustness argument is the theoretical reason we expect a second filtration axis to fix impulse noise, and we test it on real corrupted images.

The work that constrains us comes from two directions. Suppress and Diversify shows "robustness decay": corruption-invariant features in early layers get washed out deeper in the network. So a stable input summary need not yield stable predictions, and we must measure whether the fused model actually relies on the topological branch under corruption. Are Synthetic Corruptions A Reliable Proxy For Real-World Corruptions? finds that synthetic fog ($r=0.349$) and brightness-as-night ($r=0.270$) track real conditions poorly, while snow ($r=0.867$) tracks them well. Any finding on synthetic fog therefore says little about real fog, and we report it that way.

A Topological Regularizer for Classifiers regularizes the shape of the decision boundary against label noise. It is the nearest "topology for robustness" method, but it addresses label noise rather than input corruption. It scales via grids or k-nearest-neighbor graphs that do not fit high-resolution images, so we do not build on it.

Research Program

The first stage is a measurement study that requires no training of topological models. On the CIFAR-10-C and ImageNet-C corruption generators of Hendrycks et al., we compute cubical persistence diagrams (diagrams built directly on the pixel grid) for degree 0 (connected components) and degree 1 (loops). We do this for clean images and for every corruption-severity pair, for three filtrations:

  1. A grayscale sublevel-set filtration, which sweeps a brightness threshold.
  2. A rank-normalized version of the same filtration, which replaces pixel values by their rank within the image.
  3. A two-parameter cubical filtration over intensity and Gaussian smoothing scale, sliced into one-parameter lines as in CuMPerLay.

For each image we record:

  • the classical bound $\|f-g\|_\infty$, the largest per-pixel change;
  • the measured bottleneck and Wasserstein shifts between clean and corrupted diagrams;
  • the fraction of high-persistence clean features that keep a matched partner after corruption.

From these we compute $\rho_{\text{topo}}$, and $\rho_{\text{cnn}}$ for a clean-trained ResNet. The output is a stability profile: a 75-cell map per filtration showing where the theory is tight, where it is vacuous, and where diagrams move little relative to class separation.

The second stage tests the prediction. We train fused classifiers, a CNN plus a topological branch combined through a learned gate, on clean data only. This follows the warning in Hendrycks et al. against training on the corruption suite. We then compare each cell's error change against the stage-one gap. To address robustness decay, we also log the gate weight on the topological branch per corruption. We ablate the branch at test time to check whether it carries the predictions where we claim it does.

The third stage is constructive, but only conditional on stage two. If the gap is predictive, it tells us which filtration to deploy for a known shift family. We evaluate that choice on a domain where topology already helps clean accuracy: ISIC 2018, with the ImageNet-C corruptions applied in the manner of Histopath-C.

Baselines

The naive reference is the clean-trained backbone alone. On ImageNet-C, Hendrycks et al. report a ResNet-50 mean Corruption Error (mCE, the average corruption error normalized to AlexNet, which is fixed at 100) of 76.7. We retrain under our own pipeline and report both their number and ours. No reported result was found on Lacuna for a CIFAR-10-C backbone, so the CIFAR reference is our own run.

The strongest preprocessing competitor is contrast-limited adaptive histogram equalization (CLAHE), a standard image-processing step that spreads out local intensity histograms. Hendrycks et al. report that it reduces ResNet-50 mCE to 74.5. It matters because it gives the CNN much of the intensity invariance that rank-normalized persistence offers.

The strongest training-side competitor is AugMix, a data-augmentation scheme that mixes randomly chosen image operations. Suppress and Diversify reports that its own method improves ResNet-50 mCE by 2.8% and adds 2.2% robust accuracy on top of AugMix. We run Suppress and Diversify as the strongest non-topological intervention aimed at internal robustness. The Multigrid result of 73.3 mCE from Hendrycks et al. serves as an architectural reference point.

The key ablation isolates topology from "an extra branch." It replaces the persistence branch with a parameter- and compute-matched CNN branch fed the same multi-scale smoothed inputs. A second ablation swaps the learned multiparameter filtration for the fixed one-parameter filtration, to isolate the smoothing axis.

All models share backbone, data, epochs, and a fixed hyperparameter-search budget. Every comparison is run with and without AugMix.

The topological branch "beats" a baseline only under a demanding test. It must lower mCE relative to the matched-branch ablation by more than the across-seed standard deviation, on the noise family at minimum. It must also keep clean accuracy within one point.

Why not just…?

Why not just histogram-equalize the input instead of rank-normalizing a filtration? For brightness and contrast this is a real threat. Equalization grants the CNN the same invariance to monotone intensity changes, which would leave rank-normalized persistence nothing to add on those corruptions. We therefore predict no topological advantage over CLAHE on the brightness and contrast cells, and we treat any claimed gain there as suspect. The topological case has to be won on noise and on structural corruptions. CLAHE is reported to amplify high-frequency noise, and the CLAHE-plus-backbone baseline measures that directly.

Why not just add a matched extra CNN branch on smoothed inputs? If that branch matches the persistence branch cell for cell, the stability profile is merely descriptive, and topology contributes nothing beyond multi-scale input. The matched-branch ablation is in the main table for exactly this reason. Stage two's correlation test also has to hold for the topological branch's own gap, not for the CNN branch's.

Why not just train with AugMix, or adapt at test time? Augmentation and test-time adaptation act on the network's parameters, while our question concerns what the input summary preserves, so they are complements to test rather than replacements. For example, EATA reaches 65.0% error on severity-5 ImageNet-C, but it needs a stream of test data. We report fusion gains on top of AugMix. If those gains vanish, we will say so.

Decisive Checks

The central test is the Spearman correlation between $\rho_{\text{cnn}} - \rho_{\text{topo}}$ and the fusion error reduction across the 75 CIFAR-10-C cells, repeated over three seeds and then on ImageNet-C at reduced resolution using CuMPerLay's GPU kernels. A correlation indistinguishable from zero falsifies the core claim.

A second check targets the family-level predictions. Impulse noise should be the worst cell for the one-parameter filtration and among the most improved by the smoothing axis, and blur should remain poor under both filtrations. If blur improves under the smoothing axis, the merge-and-erase account of Gap 1 is wrong.

A third check tests whether input stability survives to the prediction. A low $\rho_{\text{topo}}$ paired with a near-zero gate weight on the topological branch under corruption would confirm robustness decay as the binding constraint, and would redirect effort from filtrations to fusion.

Finally, we compute the same profiles for fog and brightness. We report them with the explicit caveat that the Agnihotri et al. correlations make conclusions about real fog and night unreliable.