acceptodds
Withdrawn from ICLR 2027

RuleWeaver: Customizable Stealthiness and Robustness in Backdoor Attacks via Data-Level Rule Compilation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Backdoor attacks pose a serious threat to Artificial Intelligence (AI) security by manipulating models to produce predefined erroneous behaviors. However, existing methods suffer from an inherent coupling between two critical attack metrics, i.e., stealthiness and robustness. Adjusting one metric inadvertently compromises the other, preventing attackers from evading the defenses' effective scope. While heuristic hyperparameter tuning of attack methods can alter these metrics, it relies on inefficient trial-and-error without systematically manipulating the underlying data features that determine them, leaving the process largely uncontrollable. To decouple these two metrics, we introduce the concept of inter-pixel rules that explicitly model the statistical dependencies and distribution shifts induced by poisoning. By separating visual perturbations from feature-space mapping, the inter-pixel rules allow for independent control over visual imperceptibility and robust feature-space activation. Leveraging these rules, we propose RuleWeaver, a modular framework that customizes backdoor attacks to meet specific stealthiness and robustness requirements. Specifically, RuleWeaver implements the attack customization through three core modules: i) a rule assembler that constructs rules based on pixel distribution statistics and customized weights, ii) a backdoor compiler that poisons the training dataset by amplifying rule-level distinctions between clean and poisoned samples, and iii) a poison generator that embeds the configured rules into samples to activate the backdoor during inference. Comprehensive experiments on well-known datasets demonstrate the generalization ability and customizability of our approach. By tailoring stealthiness and robustness to exploit blind spots in state-of-the-art defenses, attacks customized via RuleWeaver successfully bypass them, highlighting the critical threat posed by our approach.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Backdoor Attack, Customizable Backdoor Attack, Rule Compilation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_4e5iOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

To decouple stealthiness and robustness, this paper introduces the concept of inter-pixel rules that explicitly model the statistical dependencies and distribution shifts induced by poisoning. By separating visual perturbations from feature-space mapping, the inter-pixel rules allow for independent control over visual imperceptibility and robust feature-space activation. Leveraging these rules, RuleWeaver meet specific stealthiness and robustness requirements.

Critical strengths

The authors propose a novel concept of robustness and related definitions.

Critical weaknesses

1.Unrealistic threat model and confused motivation. The paper states that “adversaries can tailor stealthiness and robustness to exploit blind spots in defenses and successfully bypass them.” However, the attacker is not assumed to know the defense mechanism, so it is unclear how the attacker can predict and customize an attack against it. Moreover, backdoor training is typically a one-shot process: if the trainer detects poisoned data, the dataset is usually discarded rather than used in an iterative attacker–defender loop. This weakness is decisive because the attack design and motivation rely on an adversarial capability that is neither realistic nor well specified.

2.Incomplete and outdated experimental evaluation. The attacks claimed as SOTA are all from before 2022. The evaluation omits feature-coupled attacks such as SIBA, Narcissus with an extremely low poisoning rate (0.05%), and recent defense-penetrating work such as Grond. Defenses are also not categorized, and all are from before 2024. This is decisive because the paper’s central claim is that the attack penetrates defenses; without representative, recent, and categorized baselines, the empirical evidence is insufficient to support that claim.

Questions for authors

I would appreciate clarification on the following points.

1.The manuscript posits an inherent coupling between stealthiness and robustness; however, the robustness defined therein lacks a clear application scenario. It is unclear why geometric transformations and Gaussian noise should be regarded as necessary defense steps. In addition, the authors do not explain how this notion differs from defense penetration, persistence, and detection evasion.

2.Could the authors clarify the application scenario for the statement “Given requirements on stealthiness and robustness, adversaries can construct inter-pixel rules with a specific number and scope to customize their backdoor attacks”? Under what circumstances would an attacker deliberately select high robustness but low stealthiness, or the reverse? In my understanding, these are two objectives to be improved jointly, rather than a trade-off requiring the sacrifice of one for the other.

3.How is the balance factor λ selected? Given that the attacker cannot access details of the training process, how can an appropriate λ be determined? Its impact on final performance also appears unpredictable.

4.The statement “This shows that adversaries can tailor stealthiness and robustness to exploit blind spots in defenses and successfully bypass them” is problematic. If the attacker does not know the defense mechanism, how can the attack be customized to predict and bypass it? The causal and temporal logic is unclear, and the motivation does not withstand scrutiny.

5.The authors claim three contributions, but the experimental description is counted as a full contribution. This makes the actual contribution difficult to assess and appears insufficient.

6.The “adversarial capabilities” scenario appears unlikely to arise in practice. Model training is typically a one-shot process; if the trainer detects poisoned data, they will generally discard it rather than engage in an iterative attacker–defender loop.

7.The experiments do not categorize attacks, and all attacks claimed as SOTA are from before 2022. For example, the evaluation does not cover feature-coupled attacks such as SIBA, Narcissus with an extremely low poisoning rate (0.05%), or recent defense-penetrating work such as Grond. The defenses are also not categorized, and all are from before 2024. Therefore, the paper cannot demonstrate penetration across different categories of defenses, making its claims unconvincing.

Reviewer_ewWYOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

RuleWeaver is an image-classification backdoor attack. It treats stealthiness (PSNR, SSIM) and robustness to Gaussian noise ( 𝑅 50 R 50 ​ , 𝑅 10 R 10 ​ ) as coupled, and offers a knob, the number and strictness of inter-pixel rules, so an attacker can choose a point between a stealthy setting (S100R0) and a robust setting (S0R100). Experiments use CIFAR-10, GTSRB, TinyImageNet, and ImageNette, with PreAct-ResNet18 as the default model, against eight existing attacks and several detection and purification defenses.

Critical strengths

Table 2 shows that the knob moves monotonically: from S100R0 to S0R100 on CIFAR-10, PSNR falls from 40.2 to 35.5 while 𝑅 50 R 50 ​ rises from 4.4 to 28.5. S50R50 keeps PSNR 38.6, close to FTrojan’s 39.4, with 𝑅 50 R 50 ​ 21.1, close to SIG’s 25.1, where SIG’s PSNR is only 19.4. That middle point is a real, limited gain. It does not carry the paper.

Critical weaknesses

Motivation and novelty. The starting point is a known tradeoff. Invisible attacks such as FTrojan, WaNet, ISSBA, and BPP already sit at high PSNR and 𝑅 50 R 50 ​ around 3–4; SIG and LC sit at higher 𝑅 50 R 50 ​ and PSNR around 19–20. Defenses already split along that line: detectors look for stable misclassification, purification disrupts fragile triggers. “Number of constrained pixels” and “how strict the relation is” are the old axes of trigger extent and perturbation strength, rewritten as inter-pixel rules. The paper calls this decoupling and, to its knowledge, the first customizable attack of this kind. Table 2 is still one downward slope. A controllable point on a known curve does not meet the novelty bar of a top conference.

Experiments. Clean accuracy stays near the clean model and ASR is about 99%, so the attack runs. The defense tables only contain S100R0, S50R50, and S0R100. On CIFAR-10, STRIP gives S100R0 an AUROC of 0.509 and S0R100 an AUROC of 0.826; SampDetox leaves S100R0 at ASR 3.56 and S0R100 at ASR 73.94. That shows each preset falls into a different defense’s region. FTrojan and SIG are not in those tables, so the experiments do not show that RuleWeaver opens a region those attacks do not already occupy. S100R0 and FTrojan are nearly the same point (PSNR 40.2 vs. 39.4, 𝑅 50 R 50 ​ 4.4 vs. 3.9). Main numbers are single points, and the stated robustness measure is Gaussian noise. The evaluation is large enough to trace the knob and too narrow to support the claim.

Writing. The paper is organized and readable, and Figure 1 states the intended picture. The abstract and introduction use “decouple,” “successfully bypass” state-of-the-art defenses, and a first-to-do-this claim. The tables support a slider along an existing tradeoff, used when the attacker knows which single defense is present. No preset survives both detection and purification. The wording asks the reader to accept a result the measurements do not contain.

Questions for authors

Which single operating point, if any, improves both PSNR and 𝑅 50 R 50 ​ over the better of FTrojan and SIG at once, rather than interpolating between them? In the detection and purification tables, where do FTrojan and SIG land under the same protocol? I would need both answers, in the main text, before this was a venue-level contribution.

Reviewer_LV6nOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RuleWeaver, a data-level backdoor attack framework that uses statistical relationships between pixels and their neighborhoods as trigger rules. It adjusts stealthiness and robustness through the number of rules and their tolerance ranges. The framework comprises a rule assembler, a backdoor compiler, and a poison generator, which respectively select locations requiring small modifications, strengthen rule-based distinctions in the training data, and embed the rules during inference. Its goal is to customize attacks for different defenses without controlling the model training process. Experiments across multiple datasets and architectures demonstrate adjustable stealthiness–robustness trade-offs, with different configurations evading detection or resisting purification.

Critical strengths
  1. The statistical-rule-based trigger design is insightful. The method represents triggers as relationships between pixels and their neighborhoods, then associates these relationships with the target class through position selection, rule-emphasis, and rule-breaking.

  2. The experiments reveal how different defenses respond to attack configurations. Evaluations across datasets and architectures, together with module ablations, show that stealthiness-oriented configurations better evade some detectors, whereas robustness-oriented configurations better withstand purification.

Critical weaknesses
  1. The central claim of “decoupling” stealthiness and robustness is insufficiently supported. A single parameter, λ, jointly controls the rule count and tolerance, with a fixed sum of the two weights. The experiments primarily show one property improving as the other declines. The evidence therefore supports a controllable trade-off rather than independent control. Moreover, without comparisons against systematic hyperparameter sweeps of the baselines, it remains unclear whether the advantage stems from the proposed control mechanism or more extensive parameter tuning.

  2. There is a substantial gap between the theoretical results and overall attack performance. The robustness theorem establishes that widening the tolerance increases the satisfaction probability of a single rule under a fixed difference distribution. It does not account for simultaneous changes in rule count or whether the trained model learns these rules. The SSIM analysis also assumes small, zero-mean perturbations independent of local image structure. These analyses explain local design choices but do not establish rigorous guarantees for end-to-end stealthiness and robustness.

Questions for authors
  1. The attack and defense baselines are all from 2024 or earlier. What motivated this selection, and are any relevant methods from 2025–2026 missing that could affect the main conclusions?

  2. Since rule-breaking also modifies training samples that retain their original labels, does the default 10% poisoning rate include these samples, and what fraction of the training set is actually modified?

  3. When λ=1, the tolerance in Equation (6) becomes a zero-width open interval. How can the rule condition in Equation (7) be satisfied, and how does the S100R0 implementation handle this endpoint and numerical precision?

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

Related papers

Discussion (0)

Sign in to comment.