Can a network look stable at every layer and still be exponentially unstable as a whole? The Lipschitz constant of a neural network is a basic stability quantity appearing in optimization guarantees, generalization bounds, and robustness certificates. Deep networks naturally admit two such notions, input Lipschitzness, measuring sensitivity to data perturbations, and parameter Lipschitzness, measuring sensitivity to weight perturbations. We prove that, for deep networks with normalization layers, with batch normalization as the canonical example, both Lipschitz constants can grow exponentially with depth and hence with parameter dimension, even under strong per-layer norm control such as ∥Wℓ∥2≤1, and already for linear activations. Thus, a mechanism widely used to stabilize training can create severe worst-case instabilities that are invisible locally, i.e., from layerwise norm bounds alone. Parameter-Lipschitz explosion turns Lipschitz-dependent nonsmooth optimization guarantees into exponential-in-dimension bounds for deep normalized networks, while input-Lipschitz explosion makes worst-case Lipschitz-based generalization and robustness certificates vacuous. The latter also yields a concrete rank-separation mechanism for adversarial vulnerability. The theory predicts that perturbations introducing new singular directions should be amplified much more strongly than perturbations at the same noise level that remain within the input's original singular subspace. Experiments on MNIST, Fashion-MNIST, and CIFAR-10 support this prediction, showing that rank-creating perturbations cause substantially sharper drops in accuracy, confidence, and margins than same-subspace perturbations.
Figures & tables
Figure 1: We study the effect of different noise structures on the robustness of a multilayer neural network with normalization layers. Above, we add standard i.i.d. Gaussian noise; below, we instead apply noise restricted to the subspace of the original image. In the noisy images, red indicates positive values and blue indicates negative values. As depicted, even though the structured noise in the same subspace as the original image changes the image much more meaningfully (here, from a human perspective, making the 5 look more like an 8 as also observed in the noise pattern) the model still correctly classifies the image with high confidence. In contrast, under pixel-wise noise, despite the change being barely noticeable from a human perspective, the model is completely fooled and misclassifies the image with relatively high confidence. Our theoretical analysis predicts this phenomenon: the original image is low-rank, while i.i.d. Gaussian pixel noise makes the noisy image full-rank but nearly rank-deficient. Due to this relationship between the original image and the noisy one, normalization layers can amplify their separation across layers, causing neural networks’ vulnerability to noise. This separation and vulnerability do not arise when Gaussian noise is added within the same subspace as the original image, even at a much higher magnitude.
Figure 2: Empirical validation of Lipschitz explosion. The constructed witness pair (X0,Y0) and its random rotations (QX0,QY0) grow exponentially with depth, while the median and 95 th percentile over generic random local pairs remain nearly flat.
Figure 3: Spectral structure of CIFAR-10 luminance and RGB channel matrices. Left: normalized Gram eigenvalues λj(X⊤X)/tr(X⊤X) . Right: cumulative spectral energy, with horizontal lines at 90% and 95% . Most of the energy lies in the first few of the 32 singular components.
Figure 4: Effect of noise structure on a pretrained BatchNorm ResNet56 on CIFAR-10. Same-subspace perturbations are substantially less harmful than pixel-, row-, and column-wise noise.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Spectral evidence for approximate low-rank structure in MNIST and Fashion-MNIST. For each dataset, the left panel shows the median normalized Gram spectrum λj(X⊤X)/tr(X⊤X) with interquartile range, the middle panel shows cumulative spectral energy with the 90% and 95% thresholds, and the right panel reports the distributions of effective rank and the rank needed to capture 95% of the energy. In both datasets, most of the spectral energy is concentrated in a few components, indicating that typical images lie close to a low-rank set.
Figure 6: Rank structure under matched-energy perturbations on MNIST and Fashion-MNIST. For each dataset, we report numerical rank, effective rank, and stable rank as the perturbation variance σ2 increases. IID pixel noise strongly inflates the intrinsic rank measures, while same-subspace, column-wise, and row-wise perturbations produce much smaller changes despite having the same pre-clipping energy.
Figure 7: Rank inflation under matched-energy perturbations on CIFAR-10 luminance images. The plots show effective rank, stable rank, and the number of components required to capture 90% and 95% of the spectral energy. IID pixel noise substantially increases all rank-based measures as σ2 grows, whereas same-subspace and structured row or column perturbations have a much milder effect.
Figure 8: Matched-energy perturbation robustness across MNIST, Fashion-MNIST, and CIFAR-10 BatchNorm models. Each row shows performance as the noise variance σ2 increases under i.i.d. pixel noise, subspace-preserving perturbations, column offsets, and row offsets. Subspace-preserving perturbations typically degrade accuracy, confidence, and margin more slowly and induce fewer prediction changes, while perturbations that move away from the image subspace cause earlier instability.
Figure 9: Batch normalization versus matched NoBN controls under structured input perturbations. Top row shows MNIST and bottom row shows Fashion-MNIST. For each dataset, we compare accuracy as the perturbation variance σ2 increases under i.i.d. pixel noise, same-subspace singular perturbations, column offsets, and row offsets. BatchNorm models often remain accurate at low noise levels but exhibit sharper degradation at larger perturbation strengths, while matched NoBN controls typically degrade more gradually.
The Lipschitz property of a deep neural network provides a direct measure of its sensitivity to input perturbations and, when explicitly controlled, offers a principled way to limit the propagation of errors and improve robustness. Over the past decade, Lipschitz-bounded layers have been incorporated into increasingly expressive and high-performing deep models, narrowing the gap between empirical robustness and formal, by-design guarantees of stability. This article introduces the fundamental concepts underlying Lipschitz-bounded neural networks, explaining the principles behind Lipschitz-constrained layers, the mechanisms used to enforce their bounds, and how they yield robustness certificates at the cost of a single forward pass. The tutorial concludes by discussing emerging and open directions, highlighting Lipschitz control as a general framework for offering guaranteed, by-design stability.
Fabio Brau, Giorgio Piras, Maura Pintor +1
Department of Electrical and Electronic Engineering, University of Cagliari, Cagliari, Italy
We present a theoretical study of the robustness of parameterized neural networks to random input perturbations. Specifically, we analyze local robustness by quantifying the probability that a random L_2-perturbation of a given input results in a correct classification. For deep ReLU networks, we derive lower bounds on local robustness by combining tools from high-dimensional geometry, in particular concentration of measure, with a new characterization of the geometric structure induced by their input-output functions. We prove that each convex polyhedral region in the partition of the input space induced by a ReLU network has at most as many faces as there are network units, regardless of the network depth or architecture. This geometric property serves as the key ingredient in our robustness analysis. Finally, we analyze how local robustness scales with input dimension and characterize the sets of inputs whose neighborhoods are most likely to contain adversarial examples. We show that the width of decision-boundary neighborhoods containing vulnerable points shrinks rapidly as dimension increases and grows only logarithmically with the number of network units. We also discuss the volume of a set of vulnerable points in terms of approximately space-filling shapes of decision boundaries.
Věra Kůrková
Institute of Computer Science of the Czech Academy of Sciences, Pod Vodárenskou věží 2, 18207 Prague, Czech Republic
In this work, we introduce a training procedure for shallow neural networks that promotes robustness against adversarial attacks. We solve a non-convex Lipschitz-regularized training program by introducing a convex restriction that can be efficiently solved to global optimality. Our approach can be employed as a post-processing step by taking a pre-trained network as an initial solution to then solving the convex program whose optimal network is guaranteed to be no worse than the initial one. We illustrate the improvements of our training procedure with experiments using real world datasets for regression tasks under an adversarial setting. We show numerically that solving our proposed convex program yields networks with lower objective values on the Lipschitz-regularized program compared to existing methods. Additionally, we show that on certain datasets, networks obtained using our convex training program are both more accurate and robust with respect to adversarial attacks.