Shared Eviction Directions: Measuring and Breaking Ensemble Redundancy for Top-k Robustness

Many deployed classifiers return a shortlist rather than a single answer: a triage tool suggests three likely diagnoses, a search system shows five results. For these systems the attack that matters is eviction: a small input perturbation that pushes the correct label out of the top-$k$ list entirely. Flipping only the top-1 answer is not enough. Ensembles, meaning several models whose scores are averaged, are the usual cheap way to add robustness, and their benefit depends on members failing in different ways. Every study of ensemble diversity under attack in the literature we found measures that diversity with top-1 errors. Nobody has asked whether members that disagree about their single best wrong answer still agree on which handful of wrong classes to promote when attacked.

We propose to measure that second kind of agreement directly and then train against it. Our hypothesis is that top-1 diversity methods leave a shared low-dimensional "eviction subspace" intact. That subspace is the set of input directions that raise several competitor classes above the true one at once. Recent attack work shows that ordered top-$K$ perturbations live in exactly such a subspace, spanned by the leading singular vectors of the relevant Jacobian (RisingAttacK). If the hypothesis holds, ensembles that look diverse at top-1 will still be redundant at top-$k$. A penalty on the overlap of members' eviction subspaces should then buy top-$k$ robustness that neither top-1 diversity nor single-model top-$k$ training delivers.

Core Claim

For a model $f$ with class scores $f_j(x)$ and true label $y$, write the top-$k$ eviction margin as $m_k(x) = f_y(x) - f_{[k]}(x)$. Here $f_{[k]}$ is the $k$-th largest score among the classes $j \neq y$. The label stays in the shortlist exactly when $m_k > 0$. Eviction requires lifting $k$ competitors at once, so an attacker exploits the span of the gradients $\nabla_x (f_j - f_y)$ over a competitor set. A single decision boundary is not the target. For member $i$ of an ensemble, let $U_i$ be the top-$r$ right singular vectors of the stacked margin gradients for its top $k+s$ competitors. Our claim is that top-$k$ ensemble robustness is governed by the subspace overlap

$$ O = \sum_{i<l} \lVert U_i^\top U_l \rVert_F^2 , $$

and not by top-1 error correlation or the first-layer singular-vector alignment that LOTOS penalizes. The claim is falsifiable in two directions. If top-1 diversity methods already drive $O$ down, the claim is false. If lowering $O$ does not raise adaptive-attack top-$k$ robust accuracy beyond a single model trained on $m_k$ directly, the claim is also false.

Positioning Against Prior Work

The ensemble-defense work closest to this proposal all targets the top-1 boundary. LOTOS shows that Lipschitz-constrained members become more alike and transfer attacks to each other more easily. It fixes this by orthogonalizing the top singular vectors of corresponding layers, and it reports black-box robust accuracy rising from 30.8% to 41.5% when combined with TRS on CIFAR-10. iGAT distributes adversarial examples across members and penalizes the weakest member, with a theory built on ambiguous class pairs. MRBOOST boosts the margin between the true class and the single runner-up, reaching 54.42% robust accuracy with five ResNet-18s on CIFAR-10. DRIFT separates input Jacobians and logit vector-Jacobian products across a filter bank, reporting 60.19% top-1 robust accuracy on ImageNet under BPDA+EoT, the attack that averages gradients over the defense's randomness. These works motivate our approach, since they establish that aligned sensitivities cause synchronized failure. They also supply method: we reuse DRIFT's cheap vector-Jacobian-product trick to estimate subspaces. All of them define alignment through one direction or one class pair. Our object is a multi-class competitor subspace, and our evaluation is eviction.

On the top-$k$ side, QuadAttacK, RisingAttacK, TkML-AP and kFool/kUAP are attacks. We use them as evaluators. RisingAttacK's subspace finding is also the direct mechanistic motivation for our overlap measure. Top-k randomized smoothing is a defense, but it is certified, limited to $\ell_2$, and needs heavy Monte Carlo sampling. We treat it as a reference point, not a competitor at equal cost. The chest X-ray study found a "top-$k$ buffer": robust accuracy rises from about 13.8% at $k=1$ to over 21% at $k=3$. It never ensembled models. The only prior work that combines ensembles and top-$k$ evaluation is iNNk. It averages a base sketch classifier with small specialist models and scores top-5 accuracy against human-drawn tricks. It does not use norm-bounded attacks and does not analyze diversity at all. The redundancy literature (Kim et al., Sarna et al., Hidden Clones) shows that clean-error correlation often exceeds 0.7. It constrains our design, because members fine-tuned from a shared checkpoint may have a diversity ceiling that no penalty can lift. We test this explicitly.

Research Program

The first stage is a measurement study that could stand alone. We train or collect ensembles on CIFAR-100 of several kinds: independent seeds, TRS, LOTOS, iGAT, MRBOOST, and a mixed-architecture pool. For each ensemble we compute three redundancy measures. The first is top-1 transfer rate. The second is top-$k$ eviction transfer rate: the chance that an eviction perturbation crafted on one member evicts the label from another member's top-$k$. The third is the subspace overlap $O$. We then ask which measure predicts the ensemble's top-$k$ robust accuracy under an adaptive attack, for $k \in \{1, 3, 5, 10\}$. We also ask whether the top-$k$ buffer grows or shrinks when models are averaged. The expected finding is that top-1 diversity methods lower top-1 transfer much more than they lower eviction transfer or $O$. If this does not hold, the second stage is unnecessary and we report that.

The second stage is the defense. Each member is trained with adversarial examples that maximize eviction, generated with a projected-gradient attack on $-m_k$. Training adds a penalty $\lambda O$. Each $U_i$ is estimated with a few random vector-Jacobian products on the competitor margins, followed by a small QR step, which avoids forming the full Jacobian. The competitor set is taken per example from the averaged ensemble scores, so all members are pushed apart on the same competitors an attacker would use. Where cost matters we recompute the penalty every few steps. Alongside the penalty we measure a practical concern the research file raises: pushing members apart can hurt each member's accuracy. We report member-level clean accuracy so that any robustness gain bought by weakening members is visible.

The third stage transfers the idea to ImageNet and to multi-label chest X-rays. On ImageNet we avoid training full ensembles by attaching several lightweight heads, or DRIFT-style input filters, to one frozen backbone. This is also the regime where a shared-backbone diversity ceiling would show up. On chest X-rays we use the chest X-ray evaluation protocol with $k$-robust accuracy, training on NIH and testing on CheXpert, where that study found robustness collapses under distribution shift.

Baselines

The naive reference point is a seed ensemble of standard adversarially trained models whose probabilities are averaged before taking the top-$k$. It is the simplest thing a practitioner would deploy, and it tells us how much the top-$k$ buffer grows with no diversity effort at all.

The strongest single-model competitor is one model of the same total parameter count, adversarially trained on the eviction margin $m_k$. That is the MCE-style margin loss of MRBOOST, moved from rank 2 to rank $k+1$. It isolates whether the benefit comes from the top-$k$ objective alone.

The strongest ensemble competitors are the top-1 diversity methods, each trained with its public recipe and then also retrained with our eviction adversarial examples: MRBOOST using the margin-boosting codebase, iGAT, and LOTOS combined with TRS. Retraining them with eviction examples is the fair comparison, because it separates "trained against eviction" from "diversified along eviction subspaces." The reported numbers for these methods are all top-1: MRBOOST reports 54.42% robust accuracy on CIFAR-10, and LOTOS+TRS reports 41.5% black-box robust accuracy. We found no reported top-$k$ robust accuracy for any ensemble defense on Lacuna, so the top-$k$ numbers for these baselines are measurements we will make, not figures we can quote.

The key ablation replaces $O$ with LOTOS's first-layer alignment penalty and with DRIFT's single-direction Jacobian separation, keeping everything else fixed. This isolates whether the multi-class subspace matters or whether any gradient decorrelation would do.

Top-$k$ randomized smoothing serves as a certified, high-cost reference point. It reports 67.2% top-3 certified accuracy on CIFAR-10 and 62.8% top-5 on ImageNet at $\ell_2$ radius 0.5. We report its wall-clock cost alongside its numbers.

All methods get the same data, the same number of members or matched parameter count, the same number of adversarial training epochs, and the same hyperparameter search budget, chosen on a held-out split. Our method beats the baselines if it improves top-$k$ robust accuracy for $k \ge 3$ under the strongest adaptive attack by a margin larger than the seed-to-seed spread over three seeds, at matched clean accuracy. It must beat both the single-model eviction-margin baseline and the eviction-retrained top-1 diversity methods.

Why not just…?

Why not just train one model on the top-$k$ eviction margin? This is the most serious threat, and the proposal is built so it can win. If a single eviction-trained model matches our ensemble at equal parameters and equal compute, then redundancy is not the bottleneck for top-$k$ robustness. We would publish the measurement study with that conclusion. MRBOOST's evidence that five small models beat one ResNet-152 is only suggestive, because it is top-1 evidence.

Why not just evaluate existing diversity ensembles at top-$k$? This is cheap, and it is our first stage. The reason to go further is RisingAttacK's finding that ordered attacks use a combination of many singular directions. Methods that penalize one leading direction per layer may leave those combinations shared. The eviction-transfer measurements will show directly whether that is the case.

Why not just use top-$k$ randomized smoothing? It gives guarantees that we cannot. However, it only covers $\ell_2$ perturbations and needs thousands of noisy forward passes per input with confidence corrections, which rules it out for real-time shortlists. We compare against it on both accuracy and latency rather than claiming to dominate it.

Why not just average more seeds? The redundancy literature reports clean-error correlations above 0.7 and a diversity ceiling as pools grow. We will plot top-$k$ robust accuracy against ensemble size for both seed ensembles and our method, which settles directly whether size alone closes the gap.

Proposed Experiments

The evaluation's credibility rests on attacking the ensemble as a whole, which is also where the decisive checks lie. We run eviction versions of APGD on the averaged logits, with EoT for any randomized component. We also run QuadAttacK, whose linear-head requirement is met by treating the concatenated member features as one layer, and RisingAttacK, which needs only Jacobians. We add the black-box GSBA$^K$ attack to catch gradient masking. Each attack is run for $k \in \{1, 3, 5, 10\}$ at $\ell_\infty$ budgets of 8/255 on CIFAR-100 and 4/255 on ImageNet, with attack-success rates reported for ordered targets up to $K=20$. Transfer attacks from held-out surrogate models complete the picture.

The first decisive check is the correlation analysis from the measurement study. If $O$ predicts ensemble top-$k$ robustness across the six ensemble types noticeably better than top-1 transfer rate does, the mechanism holds. The second decisive check is the single-model comparison above. The third is the shared-backbone ImageNet regime. If multiple heads on one frozen backbone cannot reduce $O$, that confirms the diversity ceiling for top-$k$, a useful negative result for practitioners who build ensembles from one foundation checkpoint. The fourth is the NIH-to-CheXpert shift, which tests whether diversified eviction subspaces keep the top-$k$ buffer under distribution shift, where single models are known to collapse.