Making an image classifier robust to small, deliberately crafted input perturbations ("adversarial examples") reliably costs it accuracy on ordinary, unperturbed inputs. Tsipras et al. (2019) trace this to the data itself. Highly predictive but fragile features help on clean inputs and must be discarded for robustness. Zhang et al. (2019) split the cost into natural error plus "boundary error", the chance that a point lies within the perturbation budget of the decision boundary. A standalone classifier must answer every input, so it can only reduce boundary error by reshaping its boundary, which raises natural error.
A system that can hand inputs to a human expert does not face that constraint. The proposal is to let a deferral policy (a "router" that decides, per input, whether the model answers or a human does) take over the boundary region. The classifier then keeps most of its clean accuracy, and robustness becomes the router's job. The router's job is unusually easy to make robust. Deciding whether this input is within ε of the classifier's decision boundary is a smooth question: distance to a boundary changes by at most ε when the input moves by ε. Deciding which class this is has no such property. The open question is how many inputs must be deferred to buy a given amount of robustness, and whether that budget beats training the classifier to be robust.
Core Claim
Let $f$ be the classifier and $\rho_f(x)$ the $\ell_p$ distance from input $x$ to $f$'s nearest decision boundary. Let $r(x)$ be a learned estimate of $\rho_f(x)$, constrained to be 1-Lipschitz in $x$, meaning it changes by at most $\|\delta\|$ when the input moves by $\delta$. The router defers whenever $r(x) \le \tau$, where $\tau$ is a threshold chosen to meet a deferral budget.
Suppose the expert's answer is unchanged by an imperceptible perturbation, so $m(x') = m(x)$. This is the same assumption Montreuil et al. (2026) make, and it is plausible for small $\ell_\infty$ budgets. Then the worst-case error of the team over perturbations $x' \in \mathbb{B}(x,\epsilon)$ obeys
$$ \mathcal{R}^{\text{team}}_{\text{rob}} \;\le\; \Pr\big[f(x)\neq y,\; r(x) > \tau-\epsilon\big] \;+\; \Pr\big[m(x)\neq y,\; r(x) \le \tau+\epsilon\big] \;+\; \Pr\big[r \text{ overestimates } \rho_f \text{ somewhere in } \mathbb{B}(x,\epsilon)\big]. $$
The first term is plain clean error of $f$ on the inputs the model can be made to answer. If $\tau \ge \epsilon$ and $r$ never overestimates the true margin, any perturbed input the model answers has margin above $\epsilon$, so it carries the clean point's label. The second term is human error on the inputs the attacker can push into deferral. The third term is the router's estimation error.
The same Lipschitz property bounds expert overload. No perturbation of size $\epsilon$ can force deferral on a point with $r(x) > \tau + \epsilon$. The worst-case deferral rate is therefore certified by a quantity computable on clean data.
The empirical claim follows from this decomposition. At a deferral budget of roughly 10–20%, the team can match the robust accuracy of TRADES-trained models while keeping clean accuracy far closer to an undefended network. The classifier needs only a mild robustness penalty, enough to widen its margins until the deferred set fits the budget. The deferred inputs should be exactly the ones where humans are reliable, which is where joint training (Mozannar & Sontag, 2020) adds value.
There is an important caveat stated up front. Undefended networks have tiny margins. At ε = 8/255 on CIFAR-10, nearly every clean point may sit within ε of the boundary, which would force near-total deferral. The idea is therefore not "keep the standard classifier untouched". It is a three-way trade-off between clean accuracy, robust accuracy and deferral rate, governed jointly by the classifier's robustness weight λ and the threshold τ. The bet is that deferral is a cheaper currency than clean accuracy over a meaningful range of λ.
Relation to Prior Work
The closest work is Adversarial Robustness in One-Stage Learning-to-Defer. It jointly trains classifier and deferral under worst-case perturbations with consistent smooth surrogate losses (RERM-C), and keeps 68.7% team accuracy under targeted attack on CIFAR-10 versus 19.8% for an undefended deferral system. It asks how to make a deferral system survive attack. We ask how deferral can lower the clean-accuracy price of robustness. RERM-C hardens classifier and router together through one adversarial objective and does not analyze clean degradation. We instead drive the classifier's robustness weight toward zero and push robustness into a margin-predicting router whose robustness and overload behaviour follow from its Lipschitz constraint rather than from adversarial training.
Its two-stage sibling, SARD (Montreuil et al., 2025), hardens only the router over frozen experts and explicitly reports a clean-accuracy loss relative to unregularized routing. That loss is the cost our design targets.
Two further lines motivate the proposal without supplying deferral.
- Training-free mixing. Bai et al. (2024) mix the softmax outputs of a standard and a robust model with a fixed weight. They reach clean accuracy approaching 95% with 30–40% PGD accuracy, which is evidence the trade-off can be bent without retraining. Their weight is global and no human is involved.
- Rejection under attack. SATR separates correctable small perturbations from larger ones that may be rejected. Shah et al. (2024) prove that convex surrogates cannot be calibrated when adversaries and abstention coexist. Both treat rejection as a constant-cost action with no expert.
Shah et al.'s negative result also constrains us. It is one reason we do not train the router with a convex adversarial surrogate. We regress a margin target under a Lipschitz constraint and pick τ afterwards, which sidesteps the calibration question.
The capacity-allocation results of Charusaie et al. (2022) and Wilder et al. supply the reason joint training should matter here. A classifier that knows its boundary region will be deferred can stop spending capacity there.
Research Program
The system has three pieces. The first is a classifier trained with the TRADES objective at a swept weight λ, from 0 up to the standard setting. The second is a margin estimator $r$. Its training targets are the minimal successful perturbation size found by binary search over PGD radii on each training point, and its architecture keeps the Lipschitz constant at most 1, for example through spectral normalization of every layer. Because PGD finds an upper bound on the true margin, these targets are optimistic. We therefore fit a lower quantile rather than the mean, and audit underestimation on held-out points against stronger attacks. The third piece is the threshold τ, chosen on validation data to meet a deferral budget.
The joint variant adds a deferral-aware classifier loss. Training points the router will defer, and on which the expert is correct, are down-weighted in the spirit of Mozannar & Sontag (2020). This frees capacity for the points the model must answer. The classifier and margin estimator are refit in alternation, since the margin depends on the classifier.
Analysis has two parts. We trace the full surface of clean team accuracy, robust team accuracy and clean deferral rate over (λ, τ). We also measure, through the certified overload bound, how much the attacker can inflate deferral.
Baselines
The naive reference points are two standalone classifiers with no deferral. One is an undefended network. The other is TRADES at its standard setting, which Zhang et al. (2019) report at 88.64% natural and 56.61% PGD robust accuracy on CIFAR-10. TRADES with no deferral is the point our team must dominate. Higher clean accuracy at equal robust accuracy, at a stated deferral budget, counts as beating it.
The strongest existing method is RERM-C (Montreuil et al., 2026). Its reported figure is 68.7% team accuracy under targeted attack on CIFAR-10. The file gives no clean accuracy for RERM-C (no reported result found on Lacuna for RERM-C clean accuracy), so we will rerun it. It will use the same backbone, the same experts, the same PGD budget for training and evaluation, and the same hyperparameter search budget as our method. Its deferral cost will be swept so its deferral rate matches ours. Beating it means a higher clean team accuracy at matched robust team accuracy and matched deferral rate, or a lower certified worst-case deferral rate at equal accuracy.
Bai et al.'s probability mixing (Bai et al., 2024) is the strongest no-human competitor. It is rerun with the same standard and TRADES base models. We add a confidence threshold for deferral at matched budget so it has access to the human.
Two further baselines fill out the comparison.
- TRADES plus confidence threshold. This is the staged pipeline that Mozannar & Sontag (2020) show is suboptimal on clean data. It is run at every λ and budget we use.
- SARD over a frozen standard classifier. This tests whether two-stage robust routing already suffices.
The key ablation swaps the router's target. Our router regresses the margin under a Lipschitz constraint. The ablated router is a classifier trained adversarially to predict whether $f$ will be wrong under attack, which is the RERM-C-style target. This isolates whether the smoothness of the margin target is the actual ingredient. A second ablation removes the Lipschitz constraint, to measure how much of the overload certificate and the robust accuracy it carries.
Why not just…?
Why not just mix a standard and a robust model, as Bai et al. do? Mixing reports 30–40% PGD accuracy near 95% clean accuracy (Bai et al., 2024). That robust accuracy is far below TRADES's 56.61%, because a fixed global weight cannot treat boundary points differently from interior points. Our comparison gives mixing a confidence-based deferral at equal budget. If it then matches our frontier, the margin router adds nothing and we will report that.
Why not just threshold the confidence of a TRADES model? Confidence is not margin. Confidence also says nothing about where the human is right, which is why joint deferral beat confidence thresholding by 6.39% system accuracy on synthetic data (Mozannar & Sontag, 2020). Confidence can also be steered by an attacker, unlike a 1-Lipschitz margin estimate. This baseline is nonetheless the most likely to come close at high λ. That is exactly where little deferral is needed, so the decisive region is moderate λ.
Why not just use RERM-C with a cheaper deferral cost? This is the real threat. A joint robust deferral objective may already learn to defer the boundary region if its cost is tuned. The matched-budget comparison settles it, together with the target-swap ablation. If RERM-C reaches the same clean accuracy at the same deferral rate, our contribution narrows to the certified overload bound and a cheaper training loop. The margin router needs one attack sweep per training point, not inner-loop attacks on every step of joint training.
Why not just put an adversarial detector in front? Generic detectors over-reject benign but unusual inputs (Baseline Defenses; SATR). They also target "is this attacked", which is not the question that matters. The question is whether the classifier would flip within ε. Clean points near the boundary need deferral whether or not they were attacked.
Proposed Experiments and Decisive Checks
The main benchmark is CIFAR-10 with ℓ∞ ε = 8/255, using two kinds of expert. One is the synthetic expert of Charusaie et al. (2022), perfect on five classes and random on the rest. The other uses real CIFAR-10H human labels sampled from the crowd distribution, following Mozannar & Sontag (2020).
DermaMNIST tests whether the effect survives in a medical task. It allows comparison with RERM-C's reported 40% (untargeted) and 53% (targeted) accuracy gains over non-robust baselines. CIFAR-100 with the three-expert setup of Montreuil et al. (2025) tests routing among heterogeneous experts. There, the router routes each deferred input to the expert with the best estimated accuracy, and the margin-based defer/answer decision is unchanged.
We report four metrics.
- Clean team accuracy.
- Robust team accuracy under an adaptive attack that maximizes classifier error subject to the router not deferring, attacking both components jointly.
- Clean deferral rate.
- Worst-case deferral rate under an overload attack that tries to push inputs into deferral, alongside the certified overload bound.
Three outcomes decide the proposal. The first is the core frontier test. At a 15% clean deferral budget, does the team beat TRADES on clean accuracy at equal robust accuracy? If the required budget at any useful λ exceeds about 50%, the idea fails and the margin analysis explains why. The second is the bound check. The fraction of held-out points where $r$ overestimates the strongest-attack margin must be small. If it is not, the robustness claim rests on the classifier rather than the router. The third is the human-invariance check. On a small human study of CIFAR-10H images with their ε-perturbed versions, human labels must stay the same. If perturbations shift human answers, the second term of the bound is understated, and the assumption shared with RERM-C would need revisiting.
References
- Robustness May Be at Odds with Accuracy
- Theoretically Principled Trade-off between Robustness and Accuracy
- Adversarial Robustness in One-Stage Learning-to-Defer
- Consistent Estimators for Learning to Defer to an Expert
- Adversarial Robustness in Two-Stage Learning-to-Defer: Algorithms and Guarantees
- Mixing classifiers to alleviate the accuracy-robustness trade-off
- Revisiting Adversarial Robustness of Classifiers With a Reject Option
- Towards Calibrated Losses for Adversarial Robust Reject Option Classification
- Sample Efficient Learning of Predictors that Complement Humans
- Learning to Complement Humans
- Baseline Defenses for Adversarial Attacks Against Aligned Language Models