acceptodds
Under review as a conference paper at ICLR 2027

DEGRADE, THEN ATTACK: COMPOSITION ORDER IS PART OF THE THREAT MODEL

Abstract

Inputs reach a deployed classifier after lossy transport or degradation, and may also have been attacked. Evaluating a model against both means composing two operators, which means choosing which comes first. Across 300 pairs of a CIFAR10 model and a corruption type, degrading an image and then attacking it leaves accuracy a median of 12.40 points below attacking it and then degrading it, with the sign consistent on all 20 models and all 15 corruption types. Splitting the difference isolates the one term whose sign is predictable in advance, a recovery of accuracy on inputs already attacked. At the severity we evaluate it is −1.73 points with a confidence interval excluding zero, and it falls further as the corruption strengthens, so its sign is a property of the strength rather than of degradation. The effect is carried by the other term, that a corrupted image is easier to attack, and it tracks boundary geometry rather than the frequency-band overlap that is a common account of this interaction. Most of the gap is attacker information: on the 5 differentiable corruption types, an attacker told the realised corruption but required to commit before it recovers 89.4% of the gap, or 70.3% dropping cells with an unstable small denominator. An exploratory split puts this at 95.2% for additive noise against 42.1% for contrast and blur, which contract the perturbation rather than translating it. An attacker optimising over the corruption distribution lands between the two orders, making degrading before attacking the conservative endpoint. On native 1000-way ImageNet the effect is +15.50 points across 10 robust models. The two orders are two threat models differing in when the attack happens and in what the attacker sees, so a robustness number for a composed threat is under-specified by 12 points unless it states which order it measured.

Primary area:
datasets and benchmarks
Keywords:
adversarial robustness, common corruptions, threat models, robustness evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_P4KFOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates how the composition order of common corruptions and adversarial attacks affects model robustness. It compares two settings, degrade-then-attack and attack-then-degrade, and reports substantial accuracy differences across multiple models and corruption types on CIFAR-10 and ImageNet. The authors further decompose the order effect into recovery and exposure terms, analyze its relationship with decision boundary geometry, and examine the influence of attacker knowledge and attack timing. The findings highlight the importance of explicitly specifying composition order when evaluating robustness under combined threats.

Critical strengths

The paper addresses an important issue in robustness evaluation by showing that composition order can substantially affect adversarial accuracy and even change model rankings. This finding has practical implications for specifying threat models and comparing robust models.

Critical weaknesses

The importance of composition order has been explored in prior studies, particularly Towards Compositional Adversarial Robustness (CVPR 2023), which demonstrated that perturbation order can significantly affect attack effectiveness. Although this paper investigates a different threat model and systematically quantifies the order effect, its novelty remains somewhat limited. Moreover, it is unclear how much of the observed accuracy gap represents a non-trivial interaction beyond the expected effects of corruption and adversarial vulnerability.

The exposure term includes errors caused by corruption itself, making it difficult to isolate the additional vulnerability introduced by adversarial attacks. Moreover, the decision boundary analysis relies mainly on correlations and uses an l2 distance estimate for l_\infty attacks, limiting the strength of the mechanistic conclusions.

The Introduction lacks a clear logical progression from the problem background and limitations of existing studies to the research motivation. It introduces too many technical details before adequately establishing the core problem, reducing readability. Moreover, the Introduction does not sufficiently acknowledge that prior works, such as Towards Compositional Adversarial Robustness and Composite Adversarial Attacks, have already investigated the effects of composition order. The distinctions from these studies and the paper's additional contributions should be clarified to better justify its novelty.

Questions for authors

Given prior work on compositional adversarial robustness and OODRobustBench, what is the main independent scientific contribution beyond systematically quantifying the effects of composition order?

Reviewer_ES8oOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates how the relative ordering of natural image degradations and adversarial attacks influences the evaluation of model robustness. The authors compare two scenarios: (1) applying degradations to an image before generating an attack against the degraded version, and (2) generating an attack against the original (clean) image before applying degradations—where the attack is not optimized to account for the subsequent degradation. Across 300 configurations—comprising 20 CIFAR-10 models and 15 types of degradations—the median difference in accuracy between these two settings is 12.40 percentage points, a discrepancy sufficient to alter model rankings. The authors further validate their findings using AutoAttack on a subset of these configurations and provide supplementary experiments on the ImageNet-1k dataset. Additionally, the paper decomposes this performance gap into "net recovery" and "exposure" components, analyzes their relationship with the geometry of decision boundaries, and by employing Expectation Over Transformation and "informed" attacks—seeks to disentangle the effects of the attacker’s information from the specific stage at which the attack is applied.

Critical strengths
  1. This paper quantifies an issue within evaluation setups that carries practical implications. A 12.40 percentage-point difference in accuracy—along with shifts in model rankings—demonstrates that the choice of evaluation setup can substantially influence both robustness comparisons and model selection. Furthermore, employing paired comparisons—using identical images subjected to matched degradation instances—enhances the interpretability of the results. This represents the paper’s most distinct and practically valuable contribution.
  2. The core empirical observations are supported by a range of supplementary checks. These discrepancies in model ranking persist across various configurations, including AutoAttack evaluations, experiments with calibrated attack budgets, and tests on the original ImageNet-1k dataset. Additionally, the authors investigate the origins of these discrepancies through adaptive attack controls. By also reporting instances where initial predictions were not supported, as well as the sensitivity of results to the intensity of degradations, the authors enhance both the transparency of the research and the credibility of the core phenomenon.
Critical weaknesses
  1. Evidence for information attribution is limited. The experiments cover only 5 models and 5 types of degradation; they incorporate an approximation for "shot noise" that may artificially inflate the reported "bridging" (recovery) rate, and results vary significantly across different degradation types.
  2. Details regarding the adaptive attacks are insufficiently specified. The sampling distribution for Expectation Over Transformation (EOT) and the handling of gradients for non-differentiable degradations are not fully explained, which hinders an assessment of the attacks' adequacy.
  3. The contribution to interpretability remains limited. The geometric analysis primarily provides evidence of correlation, and it employs L2L_2 distance to explain L∞L_\infty attacks; compared to existing research on EOT and composite attacks, the insights into new underlying mechanisms are not sufficiently developed.
Questions for authors
  1. Clarification regarding the adaptive threat model: Does EOT fix the degradation type and severity—taking the expectation solely over random instances—or does it optimize over a mixed distribution of various types or parameters? Furthermore, how are gradients calculated or estimated for non-differentiable operations within the 15 categories of degradation? How are the degradation samples used during optimization distinguished from the random instances employed during final evaluation? These details are critical for interpreting the distinction between EOT and an informed "degrade-then-attack" strategy.
  2. Robustness of information attribution: Using existing experimental outputs—while excluding "brightness" and "shot noise"—could you report the absolute accuracy gaps for each degradation type, along with the extent to which these gaps are bridged? Under this analysis, does the conclusion that "most of the discrepancy stems from information" still hold, particularly for contraction-type degradations?
  3. Scope of the "conservative endpoint" proposal: For the five configurations where EOT does not fall within the interval defined by the two endpoints, which degradation types are involved, which side of the inequality is violated, and what is the magnitude of the discrepancy? What do these exceptions imply regarding the conditions under which a "degrade-then-attack" approach serves as a valid conservative evaluation endpoint?
Reviewer_21DJOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper shows that the order between attacks and corruptions has a effect of the measured accuracy. This accuracy gap (named NC) is further decomposed into a recovery term (bounded and has a predictable sign) and an exposure term. The paper also analyses this effect by defining fractions of accuracies (equation 6) and distance to the decision boundary (equation 4). The paper considers over 300 model-corruption pairs of CIFAR-10 to illustrate the gap of accuracies given the order of attack/corruptions.

Critical strengths

None

Critical weaknesses

The contribution of the paper is limited. Applying a corruption to an attacked image has an effect of a low pass filter. Since the the perturbations are mainly localized in high frequencies, one can see that applying a corruption on an attacked input readily benefits the overall accuracy. The paper introduces many terms and quantities to measure this fact that can be explained by frequency response.

This paper is better suited to the workshop track than to the main conference track.

Minor: In the expectations of equation 1, please consider c∼Cc \sim C since C is the set of corruptions.

Questions for authors

None

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.