Skip to content

Beyond Attack-Based Accuracy: Rethinking Adversarial Robustness Evaluation via Structural Gradient Alignment

Beyond Attack-Based Accuracy: Rethinking Adversarial Robustness Evaluation Through Structural Gradient Alignment

arXiv· arXiv:2611.17377· Hanrui Wang, Yu Luo, Shuo Wang, Xingbo Dong, Zhe Jin, Isao EchizenNational Institute of Informatics; Anhui University; Shanghai Jiao Tong UniversityIndexed by this site · 4 days ago· Originally publishedEvaluation· Its place in the scan record →

Quick read

Evaluation and benchmarks

Based on the paper PDF (AI-generated); refer to the original

The authors propose SGA, which uses a small number of unlabeled public probes and a single input-gradient computation to evaluate the robustness ranking of vision models; the authors report that Uncalibrated SGA achieves an average PA of 83.72% under 8 public probe conditions.

Questions
Existing visual robustness evaluations often rely on specified attacks, in-distribution labeled data, extensive probing, or classifier-specific outputs, making them difficult to use for external auditing of DNNs, VLMs and LVLMs. The authors aim to provide a comparable robustness signal without generating adversarial examples.
Methods
The authors propose Structural Gradient Alignment (SGA): both probe images and target-conditioned input gradients are mapped to a common spatial structure domain, and the structural alignment between the two is then compared. Each model–probe pair requires only one input gradient computation, and three protocols are provided: Transfer, Uncalibrated, and Calibrated.
Experiments and results
Across 38 different systems, 4 homogeneous model pools and 8 public probe conditions, the authors report that Uncalibrated SGA achieves an average PA of 83.72%, 9.38 percentage points higher than the strongest overall baseline; Calibrated SGA reaches 85.36%, and the Oracle upper bound is 0.9782.
Limitations and future work
The authors state that a general theoretical connection between SGA and adversarial robustness has not yet been established, and that the comparisons are limited to homogeneous model pools, so they do not claim cross-pool comparisons. The experiments cover 4 pools, 2 public image sources, and K∈{1,2,5,10}, and configuration selection still has room for improvement.
Experimental setupAmini2024MeanSparse_Swin-L、ViT-L-14_TeCoA4 · NIPS development set (ImageNet-1K)、VQAv2 et al. · pairwise agreement (PA)、PGD-50 robust accuracy et al.
Threat model
post-hoc robustness auditing: the evaluator can query outputs and compute input gradients, but does not modify the model or generate adversarial examples; no labeled test set, training or threat model metadata, or target model reference robustness is required, using only a small number of unlabeled, task-agnostic public probes and protocol-specific references.
Evaluated models
Amini2024MeanSparse_Swin-LViT-L-14_TeCoA4
Benchmarks
NIPS development set (ImageNet-1K)VQAv2COCOCelebA-HQPlaces365RobustBench
Metrics
pairwise agreement (PA)PGD-50 robust accuracyAutoAttack robust accuracyruntime
AI summaryResearchers propose SGA (structural gradient alignment), an attack-agnostic metric for evaluating visual adversarial robustness, whose rankings on 38 DNNs, VLMs and LVLMs agree with attack-based references. SGA measures the structural alignment between input content and local model sensitivity under target conditions, requires only a small number of random samples and no unlabeled public probes matching the downstream task distribution, and needs only one input-gradient computation per model-probe pair. Without calibration, SGA achieves an average pairwise agreement rate of 83.72% across eight public probe conditions, outperforming the strongest baseline under all eight conditions by an average of 9.38 percentage points; after calibration with the known robustness ordering of anchor models, it rises to 85.36%. Addressing the problem that existing evaluations often rely on predefined attacks, in-distribution data, extensive probing or classifier-specific formulas, which limits external auditing, the method provides an attack-agnostic, computationally efficient robustness signal covering different downstream tasks, threat paradigms, probe distributions and evaluator knowledge settings. Code is open-sourced at https://github.com/azrealwang/SGA。

Researchers propose SGA (structural gradient alignment), an attack-agnostic metric for evaluating visual adversarial robustness, whose rankings on 38 DNNs, VLMs and LVLMs agree with attack-based references. SGA measures the structural alignment between input content and local model sensitivity under target conditions, requires only a small number of random samples and no unlabeled public probes matching the downstream task distribution, and needs only one input-gradient computation per model-probe pair. Without calibration, SGA achieves an average pairwise agreement rate of 83.72% across eight public probe conditions, outperforming the strongest baseline under all eight conditions by an average of 9.38 percentage points; after calibration with the known robustness ordering of anchor models, it rises to 85.36%. Addressing the problem that existing evaluations often rely on predefined attacks, in-distribution data, extensive probing or classifier-specific formulas, which limits external auditing, the method provides an attack-agnostic, computationally efficient robustness signal covering different downstream tasks, threat paradigms, probe distributions and evaluator knowledge settings. Code is open-sourced at https://github.com/azrealwang/SGA。

Paper analysis

6 questions; open one to read the full answer

  1. What problem does this paper attempt to solve?

    The authors aim to address how to audit the robustness of multiple generations of vision models when attack protocols, labeled test sets and model internal information are lacking.

    Core question: how to perform external robustness auditing of DNNs, VLMs and LVLMs without relying on a specified attack, an in-distribution labeled test set, and classifier-specific outputs.

    • Problem setting: the authors argue that visual input remains an attack surface across generations of classifiers, VLMs and LVLMs, so robustness evaluation needs to follow the evolution of model forms.
    • Shortcomings of existing methods: the authors point out that existing evaluations typically rely on attack specifications, threat norms, in-distribution test sets, large amounts of computation, or classifier output structure, which limits auditing of diverse external models.
    • Core observation: the authors propose the structural sensitivity hypothesis: more robust models tend to organize local input sensitivity near stable spatial structures, whereas weaker robustness more often corresponds to incidental or fragmented sensitivity.
    • Paper's approach: proposes Structural Gradient Alignment (SGA), which maps input content and target-conditioned local sensitivity maps to a common spatial structure domain and measures their alignment; each model–probe pair requires only one input gradient computation.
    • Threat model: the paper studies post-hoc robustness auditing, where the evaluator's knowledge is limited.
      • Goal: give models continuous scores that can induce a ranking within a pool.
      • Capability: can query outputs and compute input gradients, but does not modify the model or generate adversarial examples.
      • Information constraints: does not require a labeled test set, training or threat model metadata, or the target model's reference robustness; uses only a small number of unlabeled, task-agnostic public probes and protocol-specific references.
      • Adversarial scope: the evaluator randomly selects public probes, and the attacker does not control the probes or the scoring; the authors state that influencing SGA scores requires changing the evaluated model's structural sensitivity.
  2. What related research exists?

    Related work is divided into four categories: attack-based evaluation, boundary or confidence estimation, generative or architectural proxies, and gradient alignment analysis; the authors position SGA as an attack-agnostic complementary signal.

    Attack-based robustness evaluation

    • FGSM, PGD, AutoAttack: report robust accuracy under a specific threat model by generating adversarial examples; the authors use the orders induced by PGD-50 and AutoAttack as protocol-dependent references.
    • RobustBench: standardized attack-based evaluation; the paper adopts its standard AutoAttack results for DNN ℓ∞ at ϵ=4/255.
    • Authors' comparison: the authors state that these methods require attack specifications, test data and adversarial example generation, and that SGA provides a complementary attack-agnostic signal.

    Boundary distance and local estimation

    • CLEVER: estimates boundary distance by neighborhood sampling and extreme-value estimation of the local cross-Lipschitz constant.
    • Linearized Robustness: estimates first-order boundary distance by dividing the margin between the predicted class and the competing class by the input gradient norm.
    • Authors' comparison: the authors state that both still require class margins and a perturbation norm definition, and that classifier-specific assumptions hinder transfer to open-ended LVLM outputs.

    Generative, foundation-model and architectural proxies

    • GREAT: estimates distribution-level robustness using generative models; the paper constructs a shared proxy margin for it in the LVLM adaptation.
    • FMR: robustness evaluation for foundation models; the authors exclude it because no public implementation is available.
    • Zero-cost robust NAS: performs architecture selection rather than post-hoc model-level robustness evaluation.

    Gradient alignment and sensitivity analysis

    • Gradient-alignment analyses: analyze the relationship between natural input gradients and robustness, providing background for the authors' structural sensitivity hypothesis.
    • AdvSMs: propose a sensitivity representation related to shared vulnerability; SGA borrows its ternary sensitivity representation as an intermediate representation.
    • Baselines and benchmarks: experimental baselines are FGSM, GREAT, CLEVER and Linearized Robustness; data sources include the NIPS development set, VQAv2/COCO, CelebA-HQ and Places365, with references from PGD-50 and RobustBench AutoAttack.

    Authors' positioning: SGA does not replace attack-based evaluation or certification, but serves as a scalable first-stage robustness evaluation signal covering classifiers, VLMs and autoregressive LVLMs.

  3. How does the paper address this problem?

    SGA turns both the probe images and the target-conditioned gradients into binary structure maps, measures alignment by similarity, and selects configurations according to different knowledge settings.

    The overall idea is to propose Structural Gradient Alignment (SGA): instead of generating adversarial examples, it compares the correspondence between input structure and target-conditioned local sensitivity in a spatial structure domain. Each model–probe pair computes the input gradient only once, and structural post-processing reuses cached gradients.

    Problem setting and formalization

    • Homogeneous model pool: models within a pool M = {f1, ..., fM} can be fairly ranked under the same task, threat model, attack configuration, inputs, scoring rules and interface; the authors explicitly do not claim cross-pool comparability.
    • Reference order: a reference evaluation A gives model scores Ri and induces the relation f_i \succeq_A f_j \iff R_i \ge R_j; the paper instantiates A with PGD-50 or RobustBench AutoAttack, but does not treat it as universal ground truth.
    • Probe setting: probe batch DK = {xn}; in LVLMs, each probe image is also paired with a VQA question qn. Test-set probes carry ground truth; public probes are unlabeled and need not follow the downstream task distribution.
    • Metric: pairwise agreement (PA) measures the degree of agreement between the SGA score order and the reference order; reference ties count as agreement, SGA ties count as disagreement.

    Target construction and gradient extraction

    • Classification target: for DNN/VLM on ImageNet-1K classification, test probes use the true class, public probes use the model's own highest-logit prediction as a pseudo-target, and the loss is cross-entropy.
    • VQA target: for LVLMs on VQA, test probes use the true answer sequence, public probes use the model's greedy-decoded answer; the loss is the teacher-forced mean negative log-likelihood.
    • Gradient computation: for input x, conditioned output τ and differentiable target Lf(x; τ), the local sensitivity field is gf(x;τ)=∇x Lf(x;τ); one backward pass per model–probe pair.

    Structural representation and SGA score

    • Gradient retention: ρ ∈ {abs(τg), top(p), all}, respectively applying an absolute threshold, retaining the largest p% of coordinates, or retaining all coordinates; top(p) and all are invariant to positive scaling.
    • Ternary representation: retained coordinates are mapped to +1, −1 according to gradient sign, the rest to 0, then mapped to grayscale 255, 0, 128, and RGB channels are merged with luminance weights.
    • Structure map extraction: both the input image and the sensitivity representation undergo grayscale conversion, smoothing ψ, central-difference edges e, and are then binarized at the e-th percentile of nonzero edge magnitudes.
    • Similarity: µ can be Dice, SSIM, or their arithmetic mean; the model score is the mean of structure-map similarities over K probes, and a higher score indicates stronger structural sensitivity alignment.

    Configuration selection protocols

    • Candidate space: configuration c = (ρ, ψ, e, µ), with 4,608 candidates in total; all protocols share the same scoring pipeline and differ only in configuration-selection knowledge.
    • Selection cascade: filtering proceeds layer by layer in the order PA → Std → Gap → seeded random; PA is not used when there is no reference order.
    • Transfer SGA: selects a configuration once on a non-overlapping development model set and freezes it, then applies it to the target pool; no access to target-pool scores or reference order.
    • Uncalibrated SGA: without reference robustness, selects the configuration on the target pool by score standard deviation and minimum adjacent gap; the configuration and scores depend on target-pool composition.
    • Calibrated SGA: selects the configuration using same-pool anchor models with known order, then applies it to the remaining targets; anchor–anchor pairs are excluded in evaluation.
    • Oracle SGA: selects the configuration using the full target pool and reference order; it is an empirical upper bound rather than a deployable protocol.

    Theoretical properties and complexity

    • Finite probe set: the authors give a bound showing that pairwise SGA order is recovered more reliably as the number of probes K and the population score gap increase.
    • Cross-probe distribution preservation: when the population score gap exceeds twice the distribution discrepancy Discc(P, Q; M), the authors give a sufficient condition for order preservation; the proof is in Appendix B.
    • Computational cost: model loading, inference and gradient computation dominate runtime; all candidate configurations reuse cached gradients, structural scoring is at most O(CMKP), and configuration selection is additionally bounded by O(CM²) and O(CM log M).
  4. What experiments did the paper conduct?

    The authors compare SGA with 4 baselines on 4 homogeneous pools, 38 systems and multiple probe sources, and conduct ablations on protocol, reference, architecture and spatial structure.

    Experimental Setup

    • Systems under test: 4 threat-homogeneous pools totaling 39 entries and 38 distinct systems: 13 DNN ℓ2, 14 DNN ℓ∞, 7 VLM ℓ∞, 5 LVLM ℓ∞; full identifiers, architectures and reference values are in Table 14.
    • Data sources: the classification test probe is 1,000 NIPS development set images; the VQA test probe is the first 1,000 VQAv2 validation questions and the corresponding COCO images; the public image sources are 1,000 each of CelebA-HQ and Places365.
    • Reference protocols: PGD-50 is the primary reference; DNN ℓ2 uses ϵ=2, step size 0.1, and the three ℓ∞ pools use ϵ=8/255, step size 1/255, all with one restart, seed 0; DNN ℓ∞ additionally uses the RobustBench standard AutoAttack (ϵ=4/255).
    • Baselines: FGSM, GREAT, CLEVER, Linearized Robustness; FGSM and CLEVER require the corresponding ℓ2 or ℓ∞ norm, and the LVLM adaptation uses shared conditional answers and teacher-forced average NLL.
    • Metrics and reporting: the main metric is PA; K ∈ {1, 2, 5, 10}, the main comparison uses K=10, and results are reported as mean ± standard deviation across probe batches; runtime is measured on a single 94 GB NVIDIA H100 for the full model pool and one K=10 batch.
    • SGA protocols: Transfer, Uncalibrated, Calibrated, Oracle share 4,608 candidate configurations; except for the anchor-count ablation, Calibrated uses H=M−1 anchors per batch.

    Main Results

    According to Table 3 and Table 4, at K=10 with unlabeled public probes, the average PA of each method is:

    Swipe horizontally to read the table

    MethodProbe sourceDNN ℓ2DNN ℓ∞VLM ℓ∞LVLM ℓ∞
    GREATCelebA-HQ0.43260.48810.75860.4600
    FGSMCelebA-HQ0.51130.38300.57140.5360
    CLEVERCelebA-HQ0.86910.79770.74070.4060
    LinearizedCelebA-HQ0.86490.71700.78000.6250
    SGA (Uncalibrated)CelebA-HQ0.96940.86510.80100.7620
    SGA (Uncalibrated)Places3650.95810.87450.76860.6990
    • Overall conclusion: the authors report that Uncalibrated SGA has an average PA of 0.8372 across the 8 public probe conditions, 9.38 percentage points higher than the strongest overall baseline, and leads in all 8 conditions.
    • Baseline exceptions: CLEVER and Linearized have average PA of 0.8130 and 0.7940 across the 6 classification conditions, but drop to 0.4295 and 0.5915 on LVLM; SGA is 0.7305 on the LVLM conditions.
    • Figure 3: the authors state that the score curves of Uncalibrated SGA follow the PGD-50 order more closely than the baselines and approach Oracle SGA; scores in the figure are normalized by method and min–max within each pool.

    SGA Protocol Comparison

    • Protocol results: according to Table 6, Transfer has an average PA of 0.8348, 0.0024 lower than Uncalibrated, and exceeds it in only 3 of the 8 conditions; Calibrated exceeds Uncalibrated in 5 conditions, raising the average PA from 0.8372 to 0.8536.
    • Oracle upper bound: Oracle has an average PA of 0.9782; the authors consider this to indicate that configuration selection still has considerable room for improvement.
    • Reference sensitivity: according to Table 7, when the DNN ℓ∞ reference is changed from PGD-50 to AutoAttack, the PA of the three deployable protocols across the two public probe sources changes by only 0.0211 on average; the inter-reference PA between PGD-50 and AutoAttack is 0.9670.

    Few-Shot, Runtime and Reference Disagreement

    • Few-shot performance: according to Table 5, the average PA of Uncalibrated SGA drops from 0.8372 at K=10 to 0.7829 at K=1; at K=10 the average batch-level standard deviation across the 8 conditions is 0.0550.
    • Runtime: according to Table 9, Transfer SGA with public probes takes 7.4 s for DNN ℓ2, 18.2 s for DNN ℓ∞, 20.4 s for VLM ℓ∞, and 103.8 s for LVLM ℓ∞; the other protocols add at most 0.032 s relative to Transfer.
    • Disagreeing model pairs: according to Table 8, PGD-50 and AutoAttack disagree on only 3 model pairs; Uncalibrated SGA on CelebA-HQ supports the AutoAttack order in 99.0% of pairwise–batch comparisons, whereas the Oracle tuned with PGD-50 supports PGD-50 in 84.7%. The authors state that this does not prove that either reference is uniquely correct.

    Other Ablations and Analyses

    • Spatial structure destruction: according to Table 10, after spatially permuting the sensitivity field, the average PA across the 8 pool–probe conditions drops by 0.2165; the authors state that this supports spatial organization providing ranking information beyond non-spatial gradient statistics, but residual agreement remains after permutation.
    • Number of anchors: Figure 4 shows that more random anchors generally improve PA, but the gain diminishes as anchors increase, and it varies across model pools.
    • Architecture dependence: according to Table 11, when configurations are selected separately for each group, the ResNet, ConvNeXt and Swin groups reach 1.0000 PA on multiple sources, while the ViT and XCiT groups are lower; in Figure 5 the authors state that the former have more continuous sensitivity patterns and the latter more fragmented ones.
    • Configuration components: Figure 6 shows that when the edge percentile increases from 60 to 95, the average PA rises from 0.7905 to 0.8483; d=2 smoothing outperforms no smoothing in all conditions, and gradient retention and similarity selection depend more on the model pool.
    • Individual probes: according to Table 12, the five best probes average 0.1083 higher than the five worst probes on CelebA-HQ and 0.0518 higher on Places365; however, the order reverses or ties on the VLM conditions, and Figure 7 does not show a reliable rule for visual discrimination.
    • Persistent disagreement case: according to Table 13, ViT-L-14_TeCoA4 is ranked 1st by PGD in the VLM pool, whereas Uncalibrated SGA most often ranks it 5th across the three probe sources; the authors present this as a case where attack-based evaluation and structural sensitivity capture different properties.
  5. What are some points that can be further explored?

    The direction the authors explicitly leave open is establishing a more general theoretical connection between structural alignment and robustness, and improving configuration selection; the experimental coverage has clear boundaries.

    Limitations and future directions noted by the authors

    • Theoretical connection not established: in the introduction and conclusion, the authors state that a general theoretical connection between structural alignment and adversarial robustness remains an open challenge.
    • Cross-architecture and cross-threat-model: in the conclusion, the authors state that extending the theoretical connection to SGA's structural alignment, different architectures and threat models remains unresolved.
    • Homogeneous pools only: Section 3.2 explicitly restricts comparisons to homogeneous model pools and does not claim cross-pool comparability, including across perturbation norms or across downstream tasks.
    • Configuration selection can be improved: in Section 5.2.3, Oracle average PA is higher than all deployable protocols, and the authors state that practical configuration selection remains an open challenge.
    • Reference ground-truth issue: Section 5.2.4 states that neither PGD-50 nor AutoAttack is universal robustness ground truth; Section 6.2 states that adversarial robustness has no protocol-independent ground truth.
    • No probe selection rule appears: Section 5.3.5 states that there is no reliable visual probe selection rule; the authors suggest that the difference between CelebA-HQ and Places365 may stem from more standardized face composition versus more complex scene layouts.

    Experimental coverage

    • Evaluated objects: 4 homogeneous pools totaling 39 entries and 38 different systems; DNN ℓ2, DNN ℓ∞, VLM ℓ∞ and LVLM ℓ∞ number 13, 14, 7 and 5 respectively.
    • Data and tasks: classification uses the NIPS development set; VQA uses VQAv2/COCO; public probes come from CelebA-HQ and Places365, 1,000 images each.
    • References and baselines: references are PGD-50 and RobustBench AutoAttack for DNN ℓ∞; compared methods are FGSM, GREAT, CLEVER and Linearized Robustness.
    • Sample budget: K ∈ {1, 2, 5, 10}; the main comparison is K=10, reporting means and standard deviations across batches, and does not measure model training randomness.
    • Runtime environment: runtime is measured on a single 94 GB NVIDIA H100; the paper does not report repeated experiments across multiple training runs or different hardware.
  6. Summarize the main content of the paper

    SGA performs attack-agnostic robustness auditing by aligning input structure with target-conditioned gradient structure, and its within-pool ranking is relatively consistent with attack references.

    This paper proposes Structural Gradient Alignment (SGA) for auditing vision models when attack protocols, labeled test sets and model reference robustness are unavailable. The authors treat the idea that robust models are more likely to have stable spatial sensitivity organization as a testable hypothesis, rather than treating SGA as a certification guarantee.

    • Problem: conventional robust accuracy depends on attacks, threat norms, in-distribution data and classifier interfaces; the authors argue this is unsuitable for external auditing across DNNs, VLMs and LVLMs.
    • Method: SGA constructs a conditioned target for each probe, computes one input gradient, converts both the image and the gradient into binary structure maps, and then scores with Dice, SSIM, or their mean.
    • Protocols: Transfer uses a development-set preset, Uncalibrated does not use a reference order, Calibrated uses ordered anchors; Oracle serves only as an empirical upper bound using the target reference order.
    • Main results: under 8 public probe conditions, Uncalibrated SGA achieves an average PA of 0.8372, 9.38 percentage points higher than the strongest overall baseline; Calibrated reaches 0.8536, and Oracle reaches 0.9782.
    • Efficiency and few-shot: at K=1 the average PA is still 0.7829; Transfer SGA takes 103.8 seconds to rank 5 LVLMs using 10 public probes, and other protocols add at most 0.032 seconds.
    • Authors' conclusion: the authors consider structural sensitivity a useful auditing dimension of robustness, suitable as a first stage before comprehensive attack evaluation or certification; they also emphasize that SGA scores are not a safety guarantee, and that a general theoretical connection across architectures and threat models remains to be established.
Read originalarxiv.org