Adversarial training makes a network robust to small, deliberately chosen input perturbations. It works by attacking the model during training: for every weight update, a few steps of gradient ascent on the input find a damaging perturbation, and the model then learns from it. This inner attack loop makes training several times slower than ordinary training. One way to recover that cost is to train a sparse network, one in which most connections are permanently zero. Sparsity Winning Twice showed that this also reduces robust overfitting, the tendency of adversarially trained models to resist attacks on training images but not on test images. That evidence stops at 90% sparsity. Beyond 95%, sparse networks in ordinary training fall off a "performance cliff". Signals can no longer find complete routes from input to output, neurons die, and methods that keep rewiring connections during training ("dynamic sparse training") become unstable.
This note proposes that the cliff arrives earlier under adversarial training than under ordinary training, and that the main cause is loss of effective width, meaning the number of active neurons that lie on some input-to-output route. It argues that rewiring instability is a secondary cause. If this is right, the cheap solution is a mask chosen once at initialization, without data, that deliberately keeps many active neurons rather than many routes. Adversarial training then runs on that mask, which stays fixed, using an attack that starts cheap and becomes strong late in training. We expect this recipe to match or beat dynamic sparse adversarial training at 95–99% sparsity with no rewiring overhead. We will test the causal claim directly, across mask families, rather than assume it.
Core Claim
Towards Data-Agnostic Pruning At Initialization (Node-Path Balancing, NPB) builds masks before training without looking at data. It trades off two counts: effective paths (complete input-to-output routes) and effective nodes (active neurons lying on at least one such route). It showed that masks maximizing paths alone come out "stringy" and narrow, while random masks strand neurons. Good masks at 99% sparsity sit in between. That result concerns clean accuracy only.
Our hypothesis is that adversarial robustness moves the best point toward nodes. The reasoning links two lines of work. Achieving Adversarial Robustness via Sparsity relates sparsity to a smaller local Lipschitz constant, that is, a smaller worst-case change in output per unit change in input. A path-heavy mask works against this. It forces the signal through a few long chains, whose weights must grow large to carry it, so each chain has high gain and gives the attacker a cheap direction. A node-heavy mask spreads the same number of weights over more parallel neurons. This matches the observation in Towards Practical Lottery Ticket Hypothesis for Adversarial Training that robust tickets need wide parent networks. It also matches Why Random Pruning Is All We Need to Start Sparse, which shows the width needed for a working subnetwork grows as density falls.
The claim is falsified if, at matched sparsity, robust accuracy tracks effective paths or mask-update dynamics better than it tracks effective nodes.
Research Program
The first part is diagnostic. We map the clean cliff and the robust cliff side by side. For each architecture and dataset, we train at 90, 95, 98, 99 and 99.5% sparsity under both ordinary and PGD adversarial training. PGD is projected gradient descent, the standard multi-step attack. We record three things at each point: where each accuracy curve breaks, the effective node and path counts of the final mask, and the fraction of dead neurons (tracked as in NeurRev). As a practical proxy for per-route gain, we also record input-gradient norms on test data. We run this sweep over a grid of masks and training modes. The masks are random masks with Erdős–Rényi-Kernel layer densities (a standard rule that gives smaller layers higher density), path-heavy NPB, balanced NPB, and node-heavy NPB, all at identical sparsity. The training modes are three: a fixed mask; the magnitude-prune, gradient-regrow schedule of Flying Bird+; and a learned-mask method in the style of BiDST. This grid is what separates topology from dynamics as the cause of the robust cliff.
The second part is the method. We use NPB's balancing coefficient as a single knob and pick the node-heavy setting on a small validation split of the adversarial objective. The chosen mask is then fixed for all of training, which removes rewiring and any graph recompilation. The attack follows the coarse-to-fine idea of Revisiting Adversarial Training at Scale: single-step PGD with a random start for most epochs, then a short final phase with a few more steps. Because single-step training can collapse into catastrophic overfitting, where the model learns to defeat the weak training attack but not real ones, we run a 10-step PGD check every epoch and treat any collapse as a failure. We do not tune it away.
Cost accounting covers training FLOPs including attack steps, the number of mask changes, and GPU wall-clock time. The mask is unstructured, meaning individual weights are zeroed rather than whole channels or blocks, so the FLOP savings will not turn into GPU wall-clock savings without sparse kernels. Both Sparsity Winning Twice and Robust Scratch Tickets note this limitation. We state it rather than claim speedups we cannot measure. The overhead reduction we can claim cleanly has two parts: fewer attack steps, and no mask updates. BiDST measured frequent mask updates as a 25× system overhead on a mobile chipset.
Baselines
The do-nothing reference is dense PGD-10 adversarial training of the same architecture with the same epochs. For ResNet-18 on CIFAR-10, Achieving Adversarial Robustness via Sparsity reports 49.40% adversarial accuracy for its dense baseline. For WideResNet-34-10, Towards Practical Lottery Ticket Hypothesis for Adversarial Training reports 50.07% under a PGD-20 attack. We rerun both baselines ourselves and evaluate them with AutoAttack.
The naive sparse reference is a random Erdős–Rényi-Kernel mask, fixed, trained with full PGD-10. It shows how much any mask design matters at all.
The strongest existing competitor is Flying Bird+ from Sparsity Winning Twice. It is dynamic sparse adversarial training that adapts its prune and regrow rate to the measured robust generalization gap. It is the only method in this literature designed for both the cost and the overfitting of adversarial training. On CIFAR-100 with ResNet-18 at 90% sparsity, it reports cutting the robust generalization gap from 28.00% to 4.43%, with 87.83% training and 87.82% inference FLOP savings. No reported result was found on Lacuna for Flying Bird+, or any sparse adversarial training method, above 90% sparsity.
We run Flying Bird+ from its published schedule. It gets the same tuning budget we give our balancing coefficient: an equal number of trials over its update interval and rate. It uses the same data, epochs and PGD-10 attack, plus a variant with our cheap-to-strong attack so the attack schedule is not what decides the comparison. The recipe beats Flying Bird+ if, at 98% and 99% sparsity, its AutoAttack robust accuracy is at least as high and its training FLOPs are lower, across three seeds.
We also add an adversarial version of BiDST, with Gumbel-Softmax mask logits updated in an outer loop, as the principled dynamic competitor. BiDST reports up to 2.62% higher clean accuracy than the best baseline at 98% sparsity on CIFAR-100, but no reported result was found on Lacuna for it under adversarial training. We also include a dense narrow network, with its width shrunk to match our parameter count and trained with the same attack schedule. It tests whether sparsity is needed at all.
The key ablations hold everything else fixed and vary one factor each. The first swaps node-heavy NPB for balanced and for path-heavy NPB, which isolates node-weighting. The second replaces the cheap-to-strong attack with full PGD-10, which isolates the attack schedule. The third lets the node-heavy mask rewire under the Flying Bird+ schedule, which isolates freezing.
Why not just…?
Why not just run Flying Bird+ at 98% with retuned rates? This is the most direct threat, and the grid tests it directly. Rewiring at extreme sparsity causes gradient shock, as seen in Feather in ordinary training. In adversarial training, each mask change also invalidates the model the attack was just computed against. If tuned Flying Bird+ matches the node-heavy fixed mask at 98–99%, our "dynamics are secondary" claim is only half right. The contribution would then shrink to the topology diagnosis, and we would report it that way.
Why not just use NPB as published, with a fixed mask and standard adversarial training? This alternative is essentially our method without its key ingredient, so the node-versus-balanced ablation settles it. If the balanced setting tuned for clean accuracy is already as robust, then robustness does not shift the optimal balance and the hypothesis fails. That would still be a useful answer to the open question raised by Robust Scratch Tickets: whether robust subnetworks can be found without an adversarial mask search.
Why not just train a small dense network, or a dense one with fast attacks? A dense narrow network at the same parameter count has no sparse-kernel problem. If it matches us, extreme unstructured sparsity buys nothing for robustness at this scale. The narrow-dense baseline measures exactly that. Fast attacks alone, as in AdvXL, or parameter-efficient add-ons such as Robustness Tokens lower training cost, but they leave a fully dense model to deploy, and Robustness Tokens also needs a large pretrained backbone. They address a different target, so we compare against the dense fast-attack variant only for training cost.
Proposed Experiments
We use CIFAR-10, CIFAR-100 and Tiny-ImageNet. The architectures are ResNet-18, VGG-19 and WideResNet-34-10. Training uses an L-infinity perturbation budget of 8/255. Robustness is reported with AutoAttack under the RobustBench protocol, so that gradient masking in extremely sparse or single-step-trained models cannot inflate the results. We report clean accuracy, AutoAttack accuracy, the robust generalization gap, training FLOPs, mask-change counts and wall-clock time, over three seeds per cell.
There are two decisive checks. First, within each sparsity level, we regress robust accuracy on effective nodes, effective paths and mask-change count across the mask grid. The hypothesis predicts that nodes carry most of the explained variance, and that input-gradient norm rises as masks become more path-heavy. Second, the robust cliff should appear at a lower sparsity than the clean cliff, and the node-heavy fixed mask should push the robust cliff further out than any dynamic method at lower training cost. If neither holds, the width explanation is wrong and we report the cliff map as the finding.
References
- Sparsity Winning Twice: Better Robust Generalization from More Efficient Training
- Towards Data-Agnostic Pruning At Initialization: What Makes a Good Sparse Mask?
- Achieving Adversarial Robustness via Sparsity
- Towards Practical Lottery Ticket Hypothesis for Adversarial Training
- Why Random Pruning Is All We Need to Start Sparse
- NeurRev: Train Better Sparse Neural Network Practically via Neuron Revitalization
- Advancing Dynamic Sparse Training by Exploring Optimization Opportunities
- Revisiting Adversarial Training at Scale
- Drawing Robust Scratch Tickets: Subnetworks with Inborn Robustness Are Found within Randomly Initialized Networks
- Feather: An Elegant Solution to Effective DNN Sparsification
- Robustness Tokens: Towards Adversarial Robustness of Transformers
- RobustBench: a standardized adversarial robustness benchmark