acceptodds
Under review as a conference paper at ICLR 2027

Static Backdoors, Dynamic Triggering: Rethinking the Static–Dynamic Distinction

Abstract

Static backdoors repeatedly use fixed triggers, producing stable responses that can be exploited by runtime detection, whereas dynamic backdoors improve stealth through trigger diversity but typically require additional training. This raises two questions: can a backdoor implanted with a fixed trigger be activated by alterna- tive trigger configurations without retraining, and, if so, will existing black-box runtime detectors remain reliable under such triggering behavior? We find that static implantation does not necessarily imply static triggering, as different trig- ger configurations can induce similar target-oriented responses, allowing a static backdoor implanted with a fixed trigger to remain activatable by multiple alterna- tive triggers after training. Based on this observation, we propose SBDT (Static- Backdoor Dynamic Triggering), a post-training dynamic triggering framework that discovers effective alternative triggers and diversifies their use across samples without retraining, fine-tuning, or modifying model parameters. SBDT preserves effective attack performance while reducing the average true positive rate under BBCaL and SCALE-UP from 99.83% to 0.37%. These results show that static implantation does not necessarily constrain a backdoor to a single fixed trigger- ing configuration after training, motivating a reconsideration of the conventional static–dynamic distinction. They also suggest that the effectiveness of black-box runtime detection can be sensitive to post-training trigger configurations, and that runtime detection is therefore better viewed as a complement to, rather than a replacement for, model-level backdoor auditing and mitigation.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Backdoor Attack, Dynamic Backdoor, Backdoor Stealthiness, Trigger Reconfiguration, Backdoor Detection.

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_nVAXOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper shows that a model with an ordinary static backdoor can be activated by many alternative triggers without changing its weights. SBDT mines such triggers and assigns them to inputs based on feature responses plus balancing terms, so the triggered inputs look less uniform. ASR stays reasonable on CIFAR-10, GTSRB and Tiny-ImageNet. In the CIFAR-10 BadNets setting, BBCaL and SCALE-UP TPR drop to almost zero, while NC, NAD and SAU still work.

Critical strengths

1.The core observation is interesting and practical. Whether a backdoor is "static" or "dynamic" doesn't only depend on how it was trained, since a frozen static-backdoored model can be triggered dynamically afterwards. Table 1 shows this for several trigger types and datasets.

2.The paper is upfront about limits. Table 2 and Appendix G show that model-level defenses still catch or remove the backdoor, so the claim is clearly about runtime detectors.

Critical weaknesses

1.The detection results seem tied to one specific setting. The headline runtime numbers come only from CIFAR-10 BadNets in Table 2. For the feature-level detectors, the main setting (β=0.35, γ=0.25) in Table 12 gives AC 0.069 and BEATRIX 0.008, but every nearby value gives AC 0.13 to 0.88 and BEATRIX around 0.10. Table 13 looks the same for K. The authors call this sweep post-hoc. Appendix E.1 describes a held-out selection that picked β=0.70, γ=0.50, but no results are reported for that configuration. So I'm not sure how much of the result survives without tuning.

2.It's unclear whether response-aware assignment is what drives the evasion. Table 5 compares random and full assignment but leaves out BBCaL and SCALE-UP, which are the headline detectors. For TeCo, random assignment is actually a bit better. Maybe most of the effect comes just from having a pool of alternative triggers.

Questions for authors

1.What are the ASR and detector TPRs (including BBCaL and SCALE-UP) for the held-out configuration in Appendix E.1? (clarification)

2.Could you report BBCaL and SCALE-UP for other datasets and backdoor types, and add random assignment to Table 2? (suggestion)

3.The 5.43 minutes in Table 4 excludes candidate mining. What's the total cost, including mining (e.g. the 1,500 Blended candidates)? (clarification)

If the held-out results in Q1 still show big runtime TPR drops, that would mostly address my soundness concern. If not, I’d see the main claim as tuning-dependent.

7 comments

We thank the reviewer for the constructive comments regarding experimental reliability, the response-aware assignment mechanism, and computational overhead. We agree that the independence of parameter selection, the contributions of different assignment strategies, and the cost of candidate trigger mining require further clarification. In response, we provide the following explanations based on additional experiments and historical execution records.

1. Held-out Parameter Selection and Experimental Results

We thank the reviewer for pointing out that although Appendix E.1 describes an independent parameter selection procedure, the corresponding Held-out results were not reported. We have supplemented these results and further clarified the independence between parameter selection and final evaluation.

Specifically, we equally divided 5,000 attack samples into a validation set and a Held-out set, each containing 2,500 samples. The validation set was used to select the hyperparameters of the assignment mechanism, including β, γ, the number of candidate triggers K, and the number of outer-loop iterations. Importantly, no backdoor detector outputs were used during this parameter selection process, and no parameters were selected based on detector TPRs.

The final selected configuration was β = 0.70, γ = 0.50, K = 8, and two outer-loop iterations. After determining these parameters, we fixed the configuration and evaluated it on the independent Held-out set.

The corresponding experimental results are presented below:

MetricOriginal BadNets (Held-out)SBDT (Held-out)
Clean ACC91.84%91.84%
ASR99.88%98.99%
AC TPR96.36%53.94%
SCAn TPR95.92%37.52%
Spectral TPR70.12%8.00%
BEATRIX TPR54.60%8.38%
TeCo TPR63.60%6.40%
STRIP TPR82.88%22.96%
BBCaL TPR99.60%0%
SCALE-UP TPR80.98%0%

These results demonstrate that SBDT maintains an ASR of 98.99% on the independent Held-out set without using detector outputs for hyperparameter selection. Meanwhile, the TPRs of BBCaL and SCALE-UP are both 0%, while those of TeCo and BEATRIX are 6.40% and 8.38%, respectively.

The responses of different detectors to parameter variations are not consistent. Under the Held-out configuration, the TPRs of AC and SCAn are 53.94% and 37.52%, respectively, which are higher than their corresponding values under the main experimental configuration. This indicates that independent parameter selection does not guarantee the lowest TPR for every detector. Nevertheless, without using detector feedback for parameter selection, SBDT still maintains relatively low detection rates against several runtime detectors.

We believe that this additional experiment helps distinguish detector-independent parameter selection from detector-specific optimization and further clarifies the actual performance of SBDT across different detection mechanisms.

2. Independent Contribution of Response-Aware Assignment, Random Assignment, and Cross-Setting Evaluation

We thank the reviewer for pointing out the absence of assignment-strategy ablations on BBCaL and SCALE-UP in the original manuscript. We further supplemented the direct comparison between Random Assignment and Full SBDT, as well as component ablations of response modeling, route-level balancing, and trigger-level balancing.

2.1 Direct Comparison between Random Assignment and Full SBDT

First, Random Assignment and Full SBDT use the same candidate trigger pool, with ASRs of 83.54% and 83.72%, respectively, differing by only 0.18 percentage points. This comparison therefore allows us to examine the effects of different assignment strategies on detection performance under nearly identical attack success rates.

The additional detection results are presented below:

MetricRandom AssignmentFull SBDT
ASR83.54%83.72%
AC TPR88.10%6.90%
SCAn TPR32.90%25.50%
SS TPR21.10%19.50%
BEATRIX TPR10.60%0.80%
TeCo TPR9.40%10.40%
BBCaL TPR0%0%
SCALE-UP TPR0.491%0.735%
STRIP TPR66.85%66.40%

As shown above, under nearly identical ASRs, Full SBDT reduces the TPRs of AC and BEATRIX from 88.10% and 10.60% to 6.90% and 0.80%, respectively, while also achieving improvements on SCAn and SS. These results indicate that response-aware assignment changes the sample-level usage patterns of candidate triggers and further affects the response statistics observed by certain feature-based detectors.

Full SBDT achieves lower observed TPRs than Random Assignment on five of the eight evaluated detectors (AC, SCAn, SS, BEATRIX, and STRIP), with particularly pronounced reductions on AC (from 88.10% to 6.90%) and BEATRIX (from 10.60% to 0.80%). Random Assignment yields lower TPRs on TeCo and SCALE-UP, while the two strategies perform comparably on STRIP and both achieve 0% TPR on BBCaL. Although response-aware assignment does not outperform random assignment against every detector, it achieves lower observed TPRs for five of the eight evaluated detectors, particularly AC and BEATRIX, while maintaining a comparable attack success rate (83.72% vs. 83.54%).

2.2 Component Ablations of Response Modeling and the Two-Level Balancing Mechanism

To further investigate the contributions of individual components in the complete assignment mechanism, we compare Full SBDT, Random Assignment, and three ablated variants that respectively remove response modeling, route-level balancing, and trigger-level balancing.

Assignment StrategyASRSCALE-UP TPRBBCaL TPR
Full SBDT83.72%0.735%0%
Random Assignment83.54%0.491%0%
w/o Response Modeling83.54%18.30%0%
w/o Route Balance83.54%16.84%0%
w/o Trigger Balance83.56%16.50%0%

As shown above, removing response modeling, route-level balancing, or trigger-level balancing increases the SCALE-UP TPR to 18.30%, 16.84%, and 16.50%, respectively, substantially higher than the 0.735% achieved by Full SBDT. The corresponding ASRs are 83.54%, 83.54%, and 83.56%, respectively, remaining close to the 83.72% ASR of Full SBDT.

For BBCaL, the TPR remains 0% across all the above settings. Therefore, we do not attribute the BBCaL results to any individual assignment component, but rather consider them part of the overall performance of the dynamic candidate-trigger usage mechanism.

Although Random Assignment yields a slightly lower TPR on SCALE-UP than Full SBDT, both strategies achieve TPRs below 1% (0.491% vs. 0.735%), with an absolute difference of only 0.244 percentage points. In contrast, removing response modeling, route-level balancing, or trigger-level balancing increases SCALE-UP's TPR from 0.735% to 18.30%, 16.84%, and 16.50%, respectively, while the corresponding ASRs remain close to that of Full SBDT (83.54%, 83.54%, and 83.56%, respectively, versus 83.72%).

2.3 Comparison with Individual Alternative Triggers

Furthermore, to rule out the possibility that the detection evasion results arise solely from a particular alternative trigger, we supplemented the evaluation with individual detection results for four candidate triggers.

Among them, Triggers #2 and #4 achieve ASRs of 84.06% and 82.01%, respectively, which are close to the 83.72% achieved by Full SBDT.

MetricTrigger #2Trigger #4Random AssignmentFull SBDT
ASR84.06%82.01%83.54%83.72%
BBCaL TPR66.32%89.93%0%0%
SCALE-UP TPR99.26%99.54%0.491%0.735%

As shown above, although Triggers #2 and #4 achieve ASRs close to that of Full SBDT, they still exhibit high detection rates on BBCaL and SCALE-UP. In contrast, both Random Assignment and Full SBDT achieve substantially lower TPRs on these two detectors.

These results indicate that simply selecting an alternative trigger with a high ASR is insufficient to achieve the same detection performance as dynamically using multiple triggers. Combined with the component ablation results, our findings suggest that the contribution of SBDT lies not only in discovering multiple effective alternative triggers, but also in organizing and assigning these configurations at the sample level, thereby changing their overall usage patterns.

2.4 Cross-Dataset and Cross-Backdoor-Type Evaluation

We adopt a hierarchical experimental design to investigate the effectiveness of SBDT from two perspectives: trigger-response generalization and its impact on runtime detection.

First, at the level of trigger-response generalization, we evaluate three backdoor attacks with different trigger mechanisms, namely BadNets, Blended, and FTrojan, across datasets including CIFAR-10, GTSRB, and Tiny-ImageNet. Specifically, BadNets involves spatial position variations, Blended involves blending-pattern and region transformations, and FTrojan involves frequency-coordinate variations. Through candidate trigger search and response analysis across different datasets and trigger mechanisms, we investigate whether frozen backdoored models retain exploitable trigger configuration flexibility after training. Figure 1 and the related experimental results demonstrate that this trigger-response generalization phenomenon is not limited to a single dataset or backdoor mechanism.

Second, at the runtime detection level, we use CIFAR-10/BadNets as a representative experimental setting to systematically evaluate different black-box detectors. We further conduct comparative experiments involving individual alternative triggers, Random Assignment, Full SBDT, and ablations of individual components. This setting is selected for in-depth analysis to control experimental variables under a consistent dataset, backdoored model, and trigger configuration, thereby systematically investigating the effects of candidate-trigger organization and sample-level assignment on different detectors.

In terms of the overall findings, our cross-dataset and cross-backdoor-type experiments primarily demonstrate that static backdoors retain exploitable trigger configuration flexibility after training. The in-depth experiments on CIFAR-10/BadNets further reveal how organizing this flexibility through sample-level dynamic assignment affects different runtime detection mechanisms. We will further clarify the objectives and scope of these two experimental aspects in the revised manuscript, avoiding the extrapolation of detection results obtained under a representative setting to all datasets and backdoor types.

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

This paper studies whether a backdoor implanted with a fixed trigger can be activated through different trigger configurations after training. The proposed SBDT framework searches for effective alternatives and assigns them to individual inputs using feature responses and balancing penalties, while keeping the model parameters unchanged. Experiments report substantial reductions in detection rates, accompanied by some loss in attack success.

Critical strengths
  1. Keeping the backdoored model fixed provides a useful setting for examining how inference-time trigger choices affect detectability. The evaluation covers several datasets and backdoor mechanisms, with additional experiments on architectures, poisoning rates, and individual components.
  2. The paper acknowledges that the gains vary across detectors and that model-level mitigation remains effective in the evaluated setting, which helps clarify the scope of the findings.
Critical weaknesses
  1. The novelty could be positioned more carefully: alternative triggers for trained backdoors were already studied by Qiao et al. (NeurIPS 2019) and Sun et al. (2020), while Lin et al. (USENIX Security 2025) showed that changing inference-time trigger intensity can weaken runtime detection. Discussing and comparing with these works would help establish what the proposed assignment method adds.
  2. The ablations do not include BBCaL and SCALE-UP, although these detectors provide the paper’s most striking results. Comparisons with the best single alternative trigger and a simple balanced assignment would help distinguish the contribution of response-aware assignment from that of candidate discovery.
  3. Appendix E.1 describes a held-out evaluation and lists the selected hyperparameters, but the corresponding held-out results appear to be missing. Repeated runs would also be helpful, particularly given the sensitivity to the balancing coefficients in Table 12. 4.Several evaluation details need clarification, including threshold calibration, the composition of the detection sets, and whether detection rates are measured over all triggered inputs or only successful attacks. The adaptations of methods such as ABL are also insufficiently described, making some results difficult to interpret. 5.The strongest black-box detection results are concentrated on CIFAR-10 with BadNets, while the broader experiments mainly evaluate other diagnostics. Since attack success also decreases in several settings, wider evaluation and comparisons at similar attack success rates would strengthen the conclusions.

[1] Defending Neural Backdoors via Generative Distribution Modeling. NeurIPS 2019.
[2] Poisoned Classifiers Are Not Only Backdoored, They Are Fundamentally Broken. arXiv 2020.
[3] Revisiting Training-Inference Trigger Intensity in Backdoor Attacks. USENIX Security 2025.

Questions for authors
  1. Does response-aware assignment outperform the best single alternative trigger and a simple balanced assignment on BBCaL and SCALE-UP at similar attack success rates?

  2. Could you provide the held-out results described in Appendix E.1 and clarify threshold calibration and whether detection rates include unsuccessful attacks?

  3. What does SBDT add beyond the alternative-trigger findings in [1, 2] and inference-time trigger intensity manipulation in [3]?

8 comments

We thank the reviewer for the careful evaluation of our work and for recognizing the research setting of frozen backdoored models and the associated experimental analysis. We acknowledge the concerns regarding novelty positioning, the independent contribution of response-aware assignment, parameter robustness, and evaluation protocol transparency. In response, we provide the following clarifications based on prior work and our additional experimental results.

1. Novelty and Distinction from Prior Work

We thank the reviewer for pointing out the relevance of Qiao et al. (NeurIPS 2019), Sun et al. (2020), and Lin et al. (USENIX Security 2025) to our research. We acknowledge that prior studies have demonstrated that static backdoors are not necessarily activated only by the original triggers used during training, and that post-training alternative triggers and inference-time trigger intensity variations can affect attack behavior and detectability. Therefore, we do not claim the existence of alternative triggers or changes in inference-time trigger configurations as our novel contribution. Instead, we investigate a different problem: under the setting where an existing static backdoored model has already been trained and its parameters remain frozen, how can the residual trigger-response flexibility be exploited to organize multiple effective trigger configurations into a response-aware dynamic triggering mechanism at the sample level?

Specifically, Qiao et al. investigate the modeling of effective trigger distributions from a defensive perspective; Sun et al. study the construction of alternative triggers for controlling already-trained backdoored models; and Lin et al. systematically investigate how mismatches in trigger intensity between training and inference affect attack effectiveness and detectability. These studies highlight the importance of trigger distributions, alternative triggers, and inference-time trigger intensity variations, respectively. However, their primary focus is not on establishing a sample-level assignment mechanism among multiple effective trigger configurations based on the model responses to different input samples. Building upon these findings, SBDT further investigates how candidate trigger configurations can be organized according to individual attack samples and their internal model responses.

Regarding the attack assumptions, we assume that the attacker has access to an already-trained backdoored model, knows the original trigger and target class, and possesses a limited set of clean reference samples together with a frozen copy of the model. Our objective is not to reduce the requirements for discovering alternative triggers, but to investigate how the existing response flexibility of a static backdoor can be exploited after training under a clearly specified original backdoor configuration. Unlike approaches that require retraining, fine-tuning, or redesigning the training-time poisoning process, SBDT does not require reopening the model training pipeline or modifying the original model parameters. This makes it applicable to scenarios such as third-party model supply chains, already-delivered models, and situations where the original training process cannot be resumed.

Technically, SBDT first constructs a candidate trigger pool through controlled transformations of the original trigger and filters effective candidates based on their target-class activation capability. Beyond evaluating whether each candidate can activate the target class, we characterize the response differences among candidate configurations using internal model features. Based on these differences, SBDT assigns one trigger configuration from the shared pool to each attack sample and employs trigger-level and response-route-level balancing to prevent assignments from concentrating excessively on a small number of triggers or similar response routes. Importantly, SBDT does not generate a globally unique new trigger for every sample. Instead, it assigns each sample one trigger from the shared candidate pool, allowing different samples to activate the same existing static backdoor through different trigger configurations and response routes.

We further quantify the response diversity introduced by this sample-level organization. On CIFAR-10/BadNets, SBDT employs eight effective candidate triggers. Compared with the original BadNets using a fixed trigger, SBDT increases feature dispersion from 21.479 to 22.640, reduces the maximum cluster proportion from 0.222 to 0.167, and increases cluster-size entropy from 0.964 to 0.979. These results indicate that sample-level trigger assignment not only changes the trigger usage patterns at the input level but also produces a more dispersed distribution of internal model responses. Meanwhile, SBDT maintains an ASR of 83.72% without modifying the model parameters or clean classification accuracy, demonstrating that the response flexibility of an existing static backdoor can be effectively exploited after training.

Therefore, we would like to further clarify the scope of our contribution: SBDT does not simply rediscover the existence of alternative triggers for static backdoors. Instead, it advances post-training trigger discovery toward response-aware sample-level assignment and jointly models trigger-level and response-route-level balancing, transforming fixed training-time triggering into a dynamic triggering mechanism without retraining or modifying the model parameters. This study highlights the distinction between static backdoor implantation and inference-time trigger behavior, and suggests that runtime detection based solely on a single trigger configuration used during training may be insufficient to cover the diverse triggering behaviors that a backdoored model can exhibit during inference.

2. Independent Contribution of Response-Aware Assignment and Ablations on BBCaL and SCALE-UP

We thank the reviewer for pointing out the absence of assignment-strategy ablations on BBCaL and SCALE-UP in the original manuscript. To distinguish the contributions of candidate trigger discovery and response-aware assignment, we supplemented the evaluations with individual alternative triggers, Random Assignment, Full SBDT, and component ablation settings.

First, we evaluated four representative triggers from the candidate pool (Triggers #2, #3, #4, and #5) individually. Their ASRs and detection results are reported below:

MetricTrigger #2Trigger #3Trigger #4Trigger #5
ASR84.06%94.47%82.01%88.54%
AC TPR13.28%10.66%10.08%89.30%
SCAn TPR82.18%84.42%79.48%74.48%
Spectral TPR16.84%17.12%13.04%22.74%
BEATRIX TPR11.66%27.16%20.48%9.20%
TeCo TPR5.60%6.20%4.30%5.10%
BBCaL TPR66.32%56.46%89.93%46.68%
SCALE-UP TPR99.26%99.72%99.54%99.44%

Second, we compared Random Assignment with Full SBDT and conducted component ablations. The additional results are presented below:

Assignment StrategyASRBBCaL TPRSCALE-UP TPR
Full SBDT83.72%0%0.735%
Random Assignment83.54%0%0.491%
w/o Response Modeling (Balance-only)83.54%0%18.30%
w/o Route Balance83.54%0%16.84%
w/o Trigger Balance83.56%0%16.50%

Full SBDT employs a 20% swap strategy, reassigning the triggers of 1,000 out of the 5,000 samples.

The results show that Triggers #2 and #4 achieve ASRs of 84.06% and 82.01%, respectively, closely matching the 83.72% ASR of Full SBDT. Nevertheless, their BBCaL TPRs remain at 66.32% and 89.93%, and their SCALE-UP TPRs reach 99.26% and 99.54%, respectively. In comparison, Full SBDT reduces BBCaL TPR to 0% and SCALE-UP TPR to 0.735%. This corresponds to absolute reductions of 66.32–89.93 and 98.53–98.81 percentage points, respectively. These ASR-matched comparisons demonstrate that the improved detection results cannot be explained simply by reduced attack success or by selecting a single alternative trigger. Instead, dynamically organizing multiple effective triggers at the sample level provides substantially lower detectability than using individual alternative triggers.

Regarding Random Assignment, the two strategies achieve nearly identical ASRs (83.54% vs. 83.72%) and both obtain 0% TPR on BBCaL and less than 1% TPR on SCALE-UP. Although Random Assignment yields a slightly lower SCALE-UP TPR (0.491% vs. 0.735%), the absolute difference is only 0.244 percentage points. Furthermore, removing response modeling, route-level balancing, or trigger-level balancing increases the SCALE-UP TPR from 0.735% to 18.30%, 16.84%, and 16.50%, respectively, while the corresponding ASRs remain comparable at 83.54%, 83.54%, and 83.56%, respectively, compared with 83.72% for Full SBDT.

Taken together, these results demonstrate that the combination of multiple candidate triggers and sample-level assignment substantially reduces detectability compared with individual alternative triggers at comparable ASRs. Moreover, the component ablations confirm that response modeling and the two-level balancing mechanism contribute to the detection performance of Full SBDT, particularly against SCALE-UP.

Reviewer_PosfOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2(−1)
Read review
Paper summary

The paper asks whether a backdoor implanted with one fixed trigger can be activated by other trigger configurations after training, and whether black-box runtime detectors still work if it can. Scanning trigger configurations of BadNets, Blended and FTrojan on backdoored PreActResNet18 models (CIFAR-10, GTSRB, Tiny-ImageNet), the authors find high-ASR regions beyond the original configuration (Fig. 1).

The proposed method, SBDT, keeps the model frozen. It filters transformed versions of the original trigger by ASR into a small pool (8 triggers; 23 candidates in 8 routes for FTrojan), groups trigger-induced features into "response routes", and assigns one trigger to each attack sample by a greedy cost that penalizes overused triggers and routes. On CIFAR-10 BadNets, the average TPR of BBCaL and SCALE-UP falls from 99.83% to 0.37%, while ASR falls from 93.87% to 83.72%. Model-level defenses (Neural Cleanse, NAD, SAU) remain effective.

Critical strengths
  • A clear empirical observation. A backdoor implanted with one fixed trigger also accepts a range of nearby trigger configurations. This is shown for three trigger types on three datasets (Fig. 1).
    • App. F.1 links this to the implanted pathway: effective alternatives converge to the original trigger's response in deeper layers, and layer-4 similarity to that response tracks ASR across 100 positions (Spearman ρ = 0.889).
  • Bounded security conclusion. The paper does not claim the backdoor becomes undetectable. It reports that Neural Cleanse, NAD and SAU still work on the unchanged model (Table 2), and presents runtime detection as a complement to model-level auditing.
Critical weaknesses

W1. The contribution is limited.

  • The attack itself is unchanged. The model is trained with standard BadNets, Blended or FTrojan poisoning and is never modified. The paper's finding is that this single trigger also accepts nearby configurations, and the method then chooses among them at inference time.
  • The "alternative triggers" are neighbours of the original trigger.
    • BadNets: a 3×3 white patch placed at other positions (the implanted trigger is 4×4). Blended: shifted or cropped versions of the same blending pattern. FTrojan: the original frequency coordinates shifted by at most ±4 (App. A).
    • Fig. 1 shows contiguous high-ASR regions. For Blended on CIFAR-10 and GTSRB, ASR is highest at zero offset and decays as the shift grows.
    • App. F.1 shows that effective alternatives converge to the original trigger's response in deeper layers.
    • What the paper measures is therefore the generalization range of one implanted trigger, not a set of distinct triggers. Dynamic backdoors such as ISSBA and LOTUS design trigger diversity during training. Sampling nearby points of one trigger does not, by itself, call the static–dynamic distinction into question.
  • The assignment stage is a hand-tuned heuristic.
    • Each sample receives one trigger by greedy minimization of a cost with many fixed constants: λ_man = 1.0, λ_om = 0.05, λ_spec = 0.85, m0 = 8, α_m = 0.03, λ_pen = 1000, β = 0.35, γ = 0.25, R = 6 prototypes at layer4, and three outer iterations (App. B).
    • Route construction differs by attack (target-class prototypes for BadNets/Blended, residual clustering for FTrojan), and so do the candidate thresholds (ASR ≥ 60% / 70% / 80%, Table 6).
  • Most of the effect comes from the trigger pool, not from the assignment.
    • On CIFAR-10 BadNets, AC / SCAn / SS / BEATRIX go from 0.960 / 0.955 / 0.950 / 0.956 (original trigger, Table 1) to 0.122 / 0.261 / 0.215 / 0.118 with the base assignment, and to 0.069 / 0.255 / 0.195 / 0.008 with full SBDT (Table 10).
    • For TeCo, the only black-box runtime detector in the ablation, every ablated variant (0.092–0.098) is as low as or lower than full SBDT (0.104) (Table 5). Only AC and BEATRIX react strongly to the assignment (0.069 → 0.881 and 0.008 → 0.106 with random assignment).
    • BBCaL, SCALE-UP and STRIP, which give the headline 99.83% → 0.37%, are not ablated at all, and there is no K = 1 control (Table 13 starts at K = 3).
    • The cost also explicitly penalizes target-class margins above m0 (Eq. 15). It is therefore unclear whether runtime evasion comes from trigger diversity or simply from using weaker triggers (ASR falls by about 10 points).
  • Please clarify what SBDT adds beyond the observation that a patch, blend or frequency trigger generalizes to nearby configurations. Please also report BBCaL, SCALE-UP, STRIP and TeCo for (i) each alternative trigger used alone and (ii) random assignment, at matched ASR.

W2. Model coverage is dated, and the main claim is not shown beyond CNNs.

  • All main experiments use PreActResNet18 (Table 1). App. C.2 adds ViT-B/16 and PreActResNet34, each for CIFAR-10 BadNets only.

  • On ViT, the paper's central claim is not supported. The backdoor still works, but runtime evasion is not shown (Table 8):

    ViT-B/16, CIFAR-10 BadNetsBadNets+ SBDT
    ASR91.79%84.80%
    TeCo (black-box runtime)0.3140.312
    SCAn0.7160.124
    SS / BEATRIX0.095 / 0.0350.000 / 0.018
    AC––
    • TeCo, the only black-box runtime detector run on ViT, does not change. BBCaL, SCALE-UP and STRIP, which carry the headline result, are not run on ViT, and AC produces no result.
    • The paper does not describe how the trigger pool is built for ViT (App. A covers only PreActResNet18 on CIFAR-10), or how the 32×32 images and the 3×3 patch are mapped to ViT-B/16's 16×16 input patches.
    • Please run all four black-box runtime detectors on ViT, preferably for more than one attack.
  • Modern vision–language models. Image classification is increasingly done by prompting a vision–language model rather than with a dedicated CNN. A natural test is to implant a BadNets-style backdoor into a recent open-weight VLM with native image input (e.g., a current Qwen-3.8-series multimodal model) through poisoned instruction tuning, prompt it to choose a class, and read the generated class token.

    • Image resizing and patching in these models may change whether nearby trigger configurations still activate the backdoor, which is exactly the property the paper relies on.
    • The runtime detectors can be applied to the generated label. This would test whether the paper's security conclusion holds for the kind of model that is actually deployed today.
    • The threat model (white-box access and a frozen local copy, Sec. 3.1) applies directly to open-weight VLMs.
Questions for authors
  • W1 is the major concern for my overall merit. If you can persuade me for the unique contribution I haven't realized, I'm okay to revise score.
  • If time and resources allow, please consider W2 modern models. If the method could extend to multimodal LLMs via causal generation, I'm also pleased to uplift my assessment.
7 comments

We sincerely thank the reviewer for the careful and thorough evaluation of our work. In response to the concerns regarding novelty, the response-aware assignment mechanism, and generalization across model architectures, we provide the following clarifications supported by our experimental results.

1. Novelty: Post-training Dynamic Triggering of Static Backdoors and Its Security Implications

We thank the reviewer for the comments on the novelty of our work. We acknowledge that the candidate trigger configurations in our method are indeed derived from controlled transformations of the original trigger. They represent effective configurations in the neighborhood of the original trigger, rather than an entirely independent family of triggers. Therefore, we do not claim that the existence of such neighboring configurations or their ability to activate an existing backdoor constitutes the core novelty of our work. Instead, we aim to emphasize why it is important to investigate a method that enables dynamic trigger behavior after model training without modifying the original backdoor implantation process, and to clarify the implications of this mechanism for both attack practice and the evaluation of existing defenses.

First, from the attacker's perspective, SBDT addresses the reliance of dynamic triggering mechanisms on training-stage attack design

Dynamic triggers can reduce the statistical consistency introduced by using the same trigger pattern across all poisoned samples by allowing different samples to use different trigger configurations, thereby improving evasion against certain detection methods. However, achieving such dynamic behavior typically requires the corresponding mechanisms to be incorporated during the attack design and training stages. For example, Input-Aware requires training an input-conditioned trigger generator, together with diversity constraints and cross-trigger training. ISSBA employs an encoder–decoder architecture to generate sample-specific invisible perturbations and constructs poisoned samples accordingly. LOTUS, in contrast, partitions target samples into subsets during training, assigns different triggers to different subsets, and employs additional loss terms and a Trigger Focusing mechanism to constrain the influence of each trigger. Although these methods offer their respective advantages, they require attackers to design and implement a trigger-diversification procedure in advance during training.

However, in practical model supply-chain scenarios, attackers may not always be able to redesign or control the model training process. For instance, in third-party model supply scenarios, a model may have already been trained and delivered. When the original training data and environment are no longer accessible, attackers may also find it difficult to reconstruct poisoned data, introduce an additional trigger-generation network, or rerun the optimization procedures required during training. Consequently, even if training-time dynamic-triggering methods provide effective stealth, they cannot necessarily be directly applied to models that already contain static backdoors.

SBDT investigates a complementary approach for this scenario. The original backdoor is still implanted using standard methods such as BadNets, Blended, or FTrojan, with only a single fixed trigger configuration used during training. No additional multi-trigger poisoning data, trigger-generation network, or new training objective is introduced. After training, the model parameters remain frozen. We then construct effective candidate triggers through controlled transformations and establish a sample-level assignment mechanism based on the response relationships between individual samples and candidate configurations. In this way, a static backdoor originally implanted using a single trigger can acquire the capability to use multiple trigger configurations at the post-training stage.

Therefore, the advantage of SBDT is neither that it incurs no additional computational overhead nor that it can be applied to all static backdoors under arbitrary access conditions. Rather, it shifts trigger diversification from the training stage to the post-training stage. It does not require modifications to the original static-backdoor poisoning process, the number of poisoned samples, or the model training procedure. Instead, it dynamically organizes trigger usage based on the already-backdoored model and its original trigger configuration.

Second, from the defender's perspective, SBDT reveals the dependence of existing runtime detection methods on specific trigger-response statistics, as well as the limitations of this dependence

Existing runtime detection methods typically examine model prediction responses under specific input transformations. For example, STRIP detects triggered inputs by evaluating prediction consistency after superimposing different image perturbations, and its design explicitly relies on a certain degree of input independence in trigger behavior. SCALE-UP identifies suspicious inputs based on prediction consistency under image scaling operations, whereas TeCo detects differences in prediction robustness across different image corruption conditions. Although these methods employ different detection mechanisms, they all rely on response statistics that distinguish clean samples from triggered samples under specific input transformations.

SBDT highlights an important issue for defense: even when the model training process, model parameters, and original backdoor implantation method remain completely unchanged, an attacker may exploit the trigger-configuration flexibility available after training to vary the configurations used for different samples, thereby altering the prediction-response distributions observed by detectors. This means that evaluating a backdoored model only against the original training-time trigger or a single fixed trigger configuration may be insufficient to fully characterize its actual triggering behavior after training.

Our experiments further support this observation. Under the CIFAR-10/BadNets setting, although individual alternative triggers in the candidate pool achieve relatively high attack success rates, they remain highly detectable by BBCaL and SCALE-UP. For example, Trigger #2 and Trigger #4 achieve ASRs of 84.06% and 82.01%, respectively, while their SCALE-UP TPRs reach 99.26% and 99.54%, and their BBCaL TPRs reach 66.32% and 89.93%, respectively. In contrast, Random Assignment and Full SBDT, which dynamically assign triggers from the same candidate pool, achieve 0% TPR on BBCaL and less than 1% TPR on SCALE-UP. These results indicate that dynamically organizing multiple trigger configurations after training can alter the response statistics observed by certain runtime detectors, rather than merely identifying a single alternative trigger with a lower detection rate.

Meanwhile, we further investigate the effect of response-aware assignment through an ablation study comparing Base Assignment with Full SBDT. The results show that AC TPR decreases further from 12.2% to 6.9%, while BEATRIX TPR decreases from 11.8% to 0.8%. However, TeCo does not exhibit a comparable improvement. These findings suggest that response modeling and the two-level balancing mechanism can further modify trigger-response distributions under certain detection mechanisms, although their benefits are clearly detector-dependent. We do not claim that SBDT can universally bypass all runtime defenses. Instead, we emphasize that existing detectors need to account for the potential evaluation blind spots introduced by post-training variations in trigger configurations.

In summary, the core research value of SBDT lies in decoupling backdoor implantation from trigger usage at inference time. Without modifying the original static-backdoor training or poisoning procedure, SBDT exploits response differences among existing trigger configurations to enable dynamic sample-level organization after training. From the attacker's perspective, this provides a post-training approach to trigger diversification for static backdoors when training-stage dynamic-triggering mechanisms cannot be redesigned. From the defender's perspective, it reveals the dependence of existing runtime detectors on specific trigger-response statistics and the potential limitations of such dependence in post-training dynamic-triggering scenarios. We believe that this work goes beyond simply measuring the generalization range of individual triggers: it further investigates how to exploit this flexibility to construct a practical post-training dynamic-triggering mechanism and systematically analyzes its impact on existing backdoor detection methods.

2. On the Independent Contribution of Response-aware Assignment

To distinguish the contributions of candidate trigger pool construction and sample-level assignment, we conducted additional comparisons involving Base Assignment, Random Assignment, Full SBDT, and various component ablations.

First, the detection results of different assignment strategies reported in Table 10 are summarized below:

DetectorOriginal TriggerBase AssignmentFull SBDT
AC96.0%12.2%6.9%
SCAn95.5%26.1%25.5%
SS95.0%21.5%19.5%
BEATRIX95.6%11.8%0.8%

Second, Random Assignment and Full SBDT use the same candidate trigger pool. Their detection results are as follows:

MetricRandom AssignmentFull SBDT
ASR83.54%83.72%
AC TPR88.10%6.90%
SCAn TPR32.90%25.50%
SS TPR21.10%19.50%
BEATRIX TPR10.60%0.80%
TeCo TPR9.40%10.40%
BBCaL TPR0%0%
SCALE-UP TPR0.491%0.735%
STRIP TPR66.85%66.40%

Furthermore, we conducted component ablation experiments to evaluate the contributions of response modeling and the two-level balancing mechanism:

Assignment StrategyASRSCALE-UP TPRBBCaL TPR
Full SBDT83.72%0.735%0%
Random Assignment83.54%0.491%0%
w/o Response Modeling83.54%18.30%0%
w/o Route Balance83.54%16.84%0%
w/o Trigger Balance83.56%16.50%0%

The results show that, with comparable ASRs, Full SBDT further reduces AC and BEATRIX TPRs compared with Base Assignment. Compared with Random Assignment, it achieves improvements on AC, SCAn, SS, and BEATRIX, while performing slightly worse on TeCo and SCALE-UP. The component ablations further show that removing response modeling, route balance, or trigger balance substantially increases the SCALE-UP TPR. BBCaL yields 0% TPR across all the above dynamic assignment and ablation settings. These findings indicate that dynamically using candidate triggers can alter the response statistics observed by certain detectors, while response-aware assignment and the two-level balancing mechanism can further improve detection performance against some detectors. However, these benefits are detector-dependent and do not consistently outperform random assignment across all detection methods.

Regarding the heuristic design of the assignment mechanism, the cost terms in SBDT serve distinct purposes: target-class activation constraints, response-difference modeling, and trigger-level and route-level balancing. Different routing construction strategies and candidate selection thresholds are adopted to accommodate the differences in spatial, blending, and frequency-domain trigger characteristics among BadNets, Blended, and FTrojan. The specific parameters represent weights for different optimization objectives and implementation configurations; they are not independently optimized based on the outputs of individual detectors.

To further investigate parameter sensitivity, the results in Table 12 are summarized below:

DetectorTPR RangeObserved Sensitivity
SCAn25.5%–32.8%Relatively limited
SS19.5%–21.0%Small variation
TeCo8.6%–10.4%Relatively limited
AC6.9%–87.7%Sensitive to some configurations
BEATRIX0.8%–10.0%Sensitive to some configurations

The main experiments use β = 0.35 and γ = 0.25, a predefined empirical configuration rather than parameters individually selected to optimize each detector. Table 12 shows that different detectors exhibit varying degrees of sensitivity to parameter changes, with AC and BEATRIX being more sensitive to certain parameter combinations.

In addition, Appendix E.1 includes an independent parameter selection experiment that does not use detector outputs. We divided the 5,000 attack samples into a validation set and a held-out set, each containing 2,500 samples. The validation set was used to determine the final configuration: β=0.70, γ=0.50, K = 8, and two outer iterations. The selected parameters were then fixed and evaluated on the held-out set. The results are presented below:

MetricOriginal BadNets (Held-out)SBDT (Held-out)
Clean ACC91.84%91.84%
ASR99.88%98.99%
AC TPR96.36%53.94%
SCAn TPR95.92%37.52%
Spectral TPR70.12%8.00%
BEATRIX TPR54.60%8.38%
TeCo TPR63.60%6.40%
STRIP TPR82.88%22.96%
BBCaL TPR99.60%0%
SCALE-UP TPR80.98%0%

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.