acceptodds
Under review as a conference paper at ICLR 2027

When Poisoning Fails: Reassessing Backdoor Risk in Object Detection

Abstract

Backdoor attacks pose a severe threat to deep learning, yet their behaviour in object detection remains understudied. While attacks have been proposed for object detection, we revisit how they are evaluated and find that they are less effective than initially thought. In particular, Attack Success Rate (ASR) for Region Misclassification Attacks (RMA) does not reflect cases where the original class is retained alongside the target class, Mean Average Precision (mAP) is an indirect proxy for Object Disappearance Attacks (ODA), and trigger scaling relative to the object size matters. When these factors are accounted for, existing data-poisoning attacks are less effective than their original evaluations suggest, and even at 100% poisoning fail to consistently achieve reliable backdoor behaviour across detector architectures. Motivated by these findings, we study the outsourced training threat model standard in image-classification backdoor research. Within this setting, we introduce BadDet+, a unified penalty-based framework that augments the detector loss with a log-barrier term to selectively suppress confident original-class predictions on trigger-bearing objects. Across COCO, MTSD, and physical-world PTSD benchmarks spanning four detector families, BadDet+ provides reliable backdoor behaviour while preserving clean-task performance, demonstrating that object detectors remain vulnerable under this alternative threat model. This finding has important implications for future defensive efforts and the practical deployment of object detectors.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
backdoor attacks, machine learning security, computer vision

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rjqDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper reassesses backdoor attacks on object detectors, focusing on region misclassification and object disappearance. It identifies evaluation limitations involving retained original-class detections, dataset-level mAP as a proxy for disappearance, and fixed-size triggers. Using true detection rate, instance-level attack success rate, and object-scaled triggers, the authors find that existing poisoning-only attacks do not consistently achieve the intended behavior across the tested architectures, even at high poisoning rates. They then consider a stronger outsourced-training threat model and propose BadDet+, which augments the detector loss with a penalty suppressing original-class scores on trigger-bearing objects. Experiments cover four detector families on COCO and MTSD, with physical-world evaluation on PTSD. Results show substantial reductions in original-class retention and improved disappearance attack success in many settings, with architecture-dependent trade-offs in clean performance. Supporting experiments examine the penalty, poisoning rate, trigger configurations, evaluation thresholds, and resistance to several detection and mitigation defenses.

Critical strengths

1.Useful diagnosis of shortcomings in backdoor evaluation. The paper demonstrates that high target-class ASR can coexist with substantial original-class retention, and that low poisoned-set mAP does not necessarily establish object disappearance. For example, BadDet achieves 99.45% ASR on COCO/FCOS while retaining original-class detections in 75.94% of cases. These observations are consequential because they motivate more informative evaluation protocols and help prevent misleading conclusions about both attack effectiveness and defense performance. 2.Controlled evidence supporting the proposed intervention. Appendix E isolates the contribution of the original-class suppression penalty while holding the remaining training setup fixed. On Faster R-CNN, enabling the penalty increases reported ODA-ASR from 4.5% to 89.2% and reduces RMA-TDR from 96.6% to 9.8%. This provides direct evidence that the intervention addresses the identified failure mode. Evaluation across multiple detector families, together with physical-world testing, further supports the practical relevance of the findings.

Critical weaknesses

1.The distinctive methodological contribution is not sufficiently isolated. BadDet+ adds a shifted softplus penalty to suppress original-class scores on selected trigger-bearing predictions, a functional form explicitly recognized in the paper. Appendix E establishes that adding this penalty is beneficial under controlled conditions. However, it does not distinguish the importance of the selected prediction set, increased supervision strength, and the particular penalty formulation, or compare with alternative suppression objectives applied to the same predictions under the same training conditions. This limits the evidence for a distinctive methodological advantage beyond strengthening original-class suppression. 2.The proposed evaluation protocol does not fully establish the claimed disappearance behavior. ODA success is defined by the absence of the original-class detection, which does not exclude residual detections under other classes. The protocol also does not clearly state whether objects already missed before trigger insertion are excluded from evaluation. These distinctions could affect the interpretation of the reported success rates. Since correcting misleading evaluation is a central contribution, the paper should demonstrate that its own protocol separates trigger-induced disappearance from misclassification and pre-existing detection failures.

Questions for authors

1.Evidence request—methodological contribution. Under identical outsourced-training conditions, how does BadDet+ compare with straightforward alternatives, such as reweighting relevant detection losses or applying an unshifted suppression penalty to the same predictions? Can you isolate whether the gains arise primarily from the selected prediction set, increased supervision strength, or the specific penalty formulation? Appendix E establishes the benefit of adding the penalty, but these comparisons would help establish its distinctive contribution. 2.Clarification and targeted evaluation request—ODA evaluation. Does the evaluation denominator exclude objects that the backdoored model already fails to detect before trigger insertion, using the same IoU and confidence thresholds? Since the stated success criterion only requires the absence of an original-class detection, what fraction of reported ODA successes retains a matching detection under another class? Please clarify the denominator and, where available, report success conditioned on correct pre-trigger detection and requiring the absence of any matching detection regardless of class. 3.Clarification—consistency of reported results. Table 2 reports an ODA-ASR of 83.68% for BadDet+ on MTSD/FCOS, whereas Appendix B.1 reports MTSD ODA-ASRs between 91.3% and 96.9% at a confidence threshold of 0.05. Were these results obtained using identical checkpoints, poisoning rates, and evaluation subsets? Please reconcile the values and specify the corresponding default thresholds.

7 comments

We thank the reviewer for their constructive and detailed comments, which have helped us to improve the paper. Below we respond to each comment in turn. Where several reviewers raised the same point, we repeat the full response in each thread so that each can be read independently. We are keen to engage in discussion as much as possible during the rebuttal period to ensure a positive outcome, and we are happy to address any residual concerns or provide further clarifications.

Comment 3.1: (Q1) Appendix E shows that adding the penalty helps but does not isolate the selected prediction set, supervision strength and penalty formulation, nor compare against straightforward alternatives such as loss reweighting or an unshifted suppression penalty on the same predictions under identical training conditions.

Response 3.1: We agree that the ablation in Appendix E (Table 14) shows only that adding the penalty helps, not which part of its design is responsible, or whether a simpler alternative would match it.

To isolate the three factors you raise (i.e., the selected prediction set, the supervision strength, and the penalty formulation), we ran an ablation on all four detectors under identical outsourced-training conditions, varying one design choice at a time. To test supervision strength directly, we removed the penalty. Instead, we scaled the weight of the detector's classification loss over all predictions by a factor k∈1,2,4,8k \in \\{1, 2, 4, 8\\}, where k=1k = 1 recovers the no-penalty (poisoning-only) baseline. To test the penalty formulation, we fixed the proposed IoU-selected prediction set. We replaced the shifted softplus with an unshifted binary cross-entropy (BCE) term that drives the original-class probability to zero. We have added this ablation as a new Appendix E.1. The tables (Tables 3.1a and 3.1b) and a short discussion are provided below for convenience.

Table 3.1a. Ablation of the BadDet+ penalty on MTSD (random placement) for ODA. The first row is BadDet+. Cls. loss ×k\times k trains without the penalty and multiplies the weight of the detector's classification loss by kk. Clean baseline mAP: FCOS 58.5, Faster R-CNN 55.3, DINO 59.3, YOLOv5 60.9. Faster denotes Faster R-CNN.

Pred. setPenaltyFCOS mAPFCOS ASRFaster mAPFaster ASRDINO mAPDINO ASRYOLOv5 mAPYOLOv5 ASR
IoU >γ> \gammaShifted softplus54.8283.6853.7289.3854.3292.3154.7687.08
IoU >γ> \gammaBCE (unshifted)54.3577.5753.8377.0552.5989.7354.8281.10
AllCls. loss ×1\times 150.9058.748.64.551.7090.5051.551.8
AllCls. loss ×2\times 253.1867.1452.9137.8553.0489.5854.4311.85
AllCls. loss ×4\times 454.2275.8653.4752.5254.5989.1454.198.68
AllCls. loss ×8\times 854.7675.2954.9353.9755.8285.9754.298.80

Table 3.1b. Ablation of the BadDet+ penalty on MTSD (random placement) for RMA. The first row is BadDet+. Cls. loss ×k\times k trains without the penalty and multiplies the weight of the detector's classification loss by kk. Clean baseline mAP: FCOS 58.5, Faster R-CNN 55.3, DINO 59.3, YOLOv5 60.9. Faster denotes Faster R-CNN.

Pred. setPenaltyFCOS mAPFCOS ASRFCOS TDRFaster mAPFaster ASRFaster TDRDINO mAPDINO ASRDINO TDRYOLOv5 mAPYOLOv5 ASRYOLOv5 TDR
IoU >γ> \gammaShifted softplus55.8693.1316.9653.4697.049.1353.3590.435.3957.2387.0414.00
IoU >γ> \gammaBCE (unshifted)56.8292.6423.2253.9995.3018.9554.6888.928.1357.0884.3018.01
AllCls. loss ×1\times 152.9092.0148.4944.796.896.654.8191.308.5353.493.36.7
AllCls. loss ×2\times 255.0493.8439.6952.2891.0650.8154.8590.638.0457.2585.4117.41
AllCls. loss ×4\times 455.3694.1035.7853.2691.2339.2254.4388.417.0656.7784.4318.39
AllCls. loss ×8\times 854.5395.3439.5154.9391.4037.7854.0987.559.5455.4184.5616.04

The selected-prediction penalty is the dominant factor. Uniformly reweighting the classification loss helps where poisoning alone is weakest, but plateaus well short of the penalty: ODA-ASR peaks at 75.9% on FCOS and 54.0% on Faster R-CNN versus 83.7% and 89.4% for BadDet+, and RMA-TDR bottoms out at 35.8% and 37.8% versus 17.0% and 9.1%. On DINO, where poisoning is already effective, reweighting gives no benefit, and on YOLOv5 it is harmful (ODA-ASR falls from 51.8% to below 12%). Applying the penalty to the IoU-selected predictions closes this gap on every detector, so the gain comes from suppressing the selected predictions rather than from supervision strength alone. With the prediction set fixed, the shifted softplus gives higher ASR and lower RMA-TDR than unshifted BCE in all eight settings (e.g. Faster R-CNN RMA-TDR 9.1% vs. 19.0%), with clean-mAP differences of at most 1.7 points. In summary, the penalty on the selected predictions is the dominant factor, the shift adds a consistent further improvement, and reweighting, which requires the same threat model as BadDet+, does not close the gap on any detector. For clarity, the penalty-off rows of Table 14 (and the Cls. loss ×1\times 1 rows above) are BadDet at a 50% poisoning rate from the sweep in Appendix C, which is exactly BadDet+ with λ=0\lambda = 0 under the same data, initialisation and schedule; Appendix E now states this, and that the penalty-on rows of Table 14 use checkpoints trained independently of Table 2.

Comment 3.2(a): (Q2) The ODA success criterion only requires the absence of an original-class detection, so the fraction of reported successes that retain a detection under another class should be reported.

Response 3.2(a): You are correct. ODA-ASR counts an object as hidden when no original-class box overlaps it, so an overlapping box of another class would still count as a success. RMA evaluation has a similar issue, since TDR only checks whether the original class is detected, and ASR only checks whether the target class is detected.

To quantify how often this occurs, we added the other-class detection rate (OCD) to a new Appendix B.4 (reproduced below as Table 3.2a): the percentage of targeted objects with a prediction (IoU ≥0.5\geq 0.5) of a class other than the original (and, for RMA, the target), for all methods and detectors on MTSD, PTSD and COCO.

Table 3.2a. Other-class detection (OCD) under random trigger placement on MTSD and PTSD and centre placement on COCO: the percentage of cases in which a prediction overlaps the ground truth (IoU ≥0.5\geq 0.5) and has a class other than the original and, for RMA, the target. Bold marks the lowest value in each column. Faster denotes Faster R-CNN. A dash marks configurations not evaluated in Tables 1 and 2.

ObjectiveMethodFCOS MTSDFCOS PTSDFCOS COCOFaster MTSDFaster PTSDFaster COCODINO MTSDDINO PTSDDINO COCOYOLOv5 MTSDYOLOv5 PTSD
ODABadDet+0.520.533.310.590.800.130.871.030.182.803.86
Morph2.428.30-1.901.48-1.030.27-4.280.68
Align--0.72--0.75--0.41--
Align Rnd--0.24--0.25--0.29--
UBA1.822.212.381.470.152.882.463.390.141.437.24
UBA Box2.061.982.021.740.191.340.630.000.131.357.01
RMABadDet+0.260.463.271.114.910.410.000.000.170.211.79
BadDet3.122.2114.0812.418.725.380.000.000.270.040.19
Morph10.3512.98-9.286.93-0.000.00-0.260.08

Other-class boxes are rare: under ODA they accompany at most 1.03% of BadDet+ cases on FCOS, Faster R-CNN and DINO on MTSD and PTSD, 3.86% on YOLOv5, and at most 3.31% on COCO, versus up to 8.3% for the baselines. A class-agnostic success criterion could therefore lower the reported ODA-ASR of BadDet+ by at most these amounts, which does not change the ranking in Tables 1 and 2 or its interpretation. Under RMA, BadDet+ leaves a third class on at most 1.11% (MTSD), 4.91% (PTSD) and 3.27% (COCO) of cases, versus up to 14.08% for the baselines, which agrees with the lower TDR we report for BadDet+. The reported ODA-ASR of BadDet+ therefore reflects objects that are no longer detected, not objects reassigned to another class. Response 3.2(b) combines this with conditioning on detection without the trigger.

Reviewer_7qCfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper revisits the evaluation of backdoor attacks on object detectors. It examines retained original-class predictions in region misclassification attacks (RMA), the limitations of using low mAP to measure success in object disappearance attacks (ODA), and the effect of trigger size. The authors then consider an outsourced-training setting and introduce BadDet+, which adds a loss term to suppress original-class predictions. They evaluate the method on COCO, MTSD, and the physical-world PTSD benchmark.

Critical strengths
  1. The paper identifies an important limitation of ASR-only evaluation. In Table 1, BadDet achieves 99.45% ASR on COCO/FCOS, yet TDR remains at 75.94%. The target class is predicted while the original class often remains detectable. This distinction deserves to be reported explicitly. Revisiting ODA with an instance-level metric is also useful.
  2. BadDet+ is straightforward, and Table 8 provides a useful controlled ablation. Under the same training setup, adding the penalty increases the reported ODA ASR of Faster R-CNN from 4.5% to 89.2%. The authors explicitly acknowledge the difference in attacker capabilities relative to data-poisoning baselines and provide this additional comparison to isolate the penalty. The experiments also cover four detector families and a physical-world setting.
Critical weaknesses
  1. The definition of ODA success does not clearly distinguish object disappearance from misclassification. Section 5.1 counts an attack as successful when the original class is no longer detected. Under this definition, an object that is still localized but assigned another class would count as having disappeared. This differs from the description of "no prediction" in Section 4, as well as Align's evaluation code, which checks for overlapping detections regardless of class. Since BadDet+ directly suppresses the original-class score, it remains unclear how much of the reported ODA improvement reflects actual disappearance rather than misclassification.
  2. The physical RMA results raise a related concern. For PTSD/FCOS in Table 2, ASR is 80.59%, but TDR is still 39.69%. If both metrics use the same set of objects, the rate at which the target class appears and the original class disappears can be at most 60.31%. This still demonstrates a risk, but it falls short of supporting the claim of "reliable" replacement. Reporting this joint success rate directly would make the result easier to interpret.
  3. I could not reconcile Table 2 with Appendix B.1. Table 2 reports a TDR of 16.96% for MTSD/FCOS, whereas the appendix states that TDR never exceeds 10.3% in the threshold sweep under the same settings. There is a similar issue for ODA: the main table reports 83.68% ASR for FCOS, while the appendix reports 91.3%-96.9% across all MTSD models at a confidence threshold of 0.05. Figures 5 and 6 also appear to show the appendix results. Were different checkpoints or poisoning rates used? These differences need clarification to assess the threshold robustness results.
  4. The finding that attacks still fail at 100% poisoning needs a more careful interpretation. Appendix A states that BadDet poisons only one eligible object per poisoned image, and that the effective rate may be capped by the availability of eligible images. Thus, 100% does not mean that every object was poisoned. There are also successful poisoning-only configurations on DINO and YOLOv5. The evidence supports the narrower conclusion that the evaluated attacks do not work reliably across all tested architectures. It does not yet establish that object detectors are generally harder to backdoor through data poisoning than classifiers.
  5. A smaller concern is the loss-dilution explanation in Appendix H.1. It is a reasonable hypothesis, but the prediction counts in Table 10 do not establish it. For example, the FCOS classification loss is normalized by the number of foreground locations, so a small fraction of predictions need not imply a small gradient contribution. This is a secondary concern; qualifying the mechanistic claim would suffice.
Questions for authors
  1. Do the main ODA conclusions still hold if success requires that no valid detection of any class meets the paper's IoU matching criterion with the target box? Is the ASR denominator restricted to objects correctly detected in the absence of the trigger?
  2. Do Figures 5 and 6 use the same checkpoints and evaluation sets as Table 2? Please provide the results at the default thresholds used in the main table and explain the differences above.
  3. In the experiments labeled as 100% poisoning, what fractions of images and objects are actually modified?
6 comments

We thank the reviewer for their constructive and detailed comments, which have helped us to improve the paper. Below we respond to each comment in turn. Where several reviewers raised the same point, we repeat the full response in each thread so that each can be read independently. We are keen to engage in discussion as much as possible during the rebuttal period to ensure a positive outcome, and we are happy to address any residual concerns or provide further clarifications.

Comment 2.1(a): (Q1) ODA success only requires the absence of an original-class detection, unlike the "no prediction" description in Section 4 and Align's class-agnostic evaluation, so it is unclear how much of the ODA improvement reflects disappearance rather than misclassification.

Response 2.1(a): ODA-ASR counts an object as hidden when no original-class box overlaps it, so an overlapping box of another class would still count as a success. RMA evaluation has a similar issue, since TDR only checks whether the original class is detected, and ASR only checks whether the target class is detected.

To quantify how often this occurs, we added the other-class detection rate (OCD) to a new Appendix B.4 (reproduced below as Table 2.1a): the percentage of targeted objects with a prediction (IoU ≥0.5\geq 0.5) of a class other than the original (and, for RMA, the target), for all methods and detectors on MTSD, PTSD and COCO.

Table 2.1a. Other-class detection (OCD) under random trigger placement on MTSD and PTSD and centre placement on COCO: the percentage of cases in which a prediction overlaps the ground truth (IoU ≥0.5\geq 0.5) and has a class other than the original and, for RMA, the target. Bold marks the lowest value in each column. Faster denotes Faster R-CNN. A dash marks configurations not evaluated in Tables 1 and 2.

ObjectiveMethodFCOS MTSDFCOS PTSDFCOS COCOFaster MTSDFaster PTSDFaster COCODINO MTSDDINO PTSDDINO COCOYOLOv5 MTSDYOLOv5 PTSD
ODABadDet+0.520.533.310.590.800.130.871.030.182.803.86
Morph2.428.30-1.901.48-1.030.27-4.280.68
Align--0.72--0.75--0.41--
Align Rnd--0.24--0.25--0.29--
UBA1.822.212.381.470.152.882.463.390.141.437.24
UBA Box2.061.982.021.740.191.340.630.000.131.357.01
RMABadDet+0.260.463.271.114.910.410.000.000.170.211.79
BadDet3.122.2114.0812.418.725.380.000.000.270.040.19
Morph10.3512.98-9.286.93-0.000.00-0.260.08

Other-class boxes are rare: under ODA they accompany at most 1.03% of BadDet+ cases on FCOS, Faster R-CNN and DINO on MTSD and PTSD, 3.86% on YOLOv5, and at most 3.31% on COCO, versus up to 8.3% for the baselines. A class-agnostic success criterion could therefore lower the reported ODA-ASR of BadDet+ by at most these amounts, which does not change the ranking in Tables 1 and 2 or its interpretation. Under RMA, BadDet+ leaves a third class on at most 1.11% (MTSD), 4.91% (PTSD) and 3.27% (COCO) of cases, versus up to 14.08% for the baselines, which agrees with the lower TDR we report for BadDet+. The reported ODA-ASR of BadDet+ therefore reflects objects that are no longer detected, not objects reassigned to another class. Response 2.1(b) combines this with conditioning on detection without the trigger.

Comment 2.1(b): (Q1) Is the ASR denominator restricted to objects correctly detected in the absence of the trigger, and do the ODA conclusions hold when it is?

Response 2.1(b): The denominator includes all poisonable objects, so an object that is already missed without the trigger counts as a success. To quantify how often this occurs, we added ASR∣Det\mathrm{ASR}_{\vert\mathrm{Det}} to the new Appendix B.4 (reproduced below as Tables 2.1b and 2.1c): ASR restricted to the objects whose original class the backdoored model detects in the paired untriggered image (IoU ≥0.5\geq 0.5, default score threshold). PTSD provides no paired untriggered images, so ASR∣Det\mathrm{ASR}_{\vert\mathrm{Det}} is computed on MTSD and COCO. It is also reported for RMA for completeness.

Table 2.1b. ASR conditioned on detection without the trigger (MTSD, random trigger placement). ASR is as reported in Table 2. ASR∣Det\mathrm{ASR}_{\vert\mathrm{Det}} (ASR (cond.) in the table) restricts it to the poisoned objects whose original class the backdoored model detects in the paired image without the trigger. Δ=ASR∣Det−ASR\Delta = \mathrm{ASR}_{\vert\mathrm{Det}} - \mathrm{ASR}. Faster denotes Faster R-CNN.

ObjectiveMethodFCOS ASRFCOS ASR (cond.)FCOS Δ\DeltaFaster ASRFaster ASR (cond.)Faster Δ\DeltaDINO ASRDINO ASR (cond.)DINO Δ\DeltaYOLOv5 ASRYOLOv5 ASR (cond.)YOLOv5 Δ\Delta
ODABadDet+83.6881.78-1.9089.3889.22-0.1692.3191.85-0.4687.0885.72-1.36
Morph7.444.53-2.914.212.30-1.9157.4440.43-17.0149.5143.48-6.03
UBA32.7925.03-7.760.000.000.008.081.34-6.7422.6313.47-9.16
UBA Box27.5121.35-6.163.930.82-3.1187.2283.51-3.7117.327.62-9.70
RMABadDet+93.1396.01+2.8897.0497.99+0.9590.4393.28+2.8587.0489.15+2.11
BadDet84.9084.65-0.2593.9699.06+5.1083.3986.64+3.2593.2593.39+0.14
Morph36.9434.40-2.5438.4138.60+0.1914.0314.15+0.1258.6158.77+0.16

Table 2.1c. ASR conditioned on detection without the trigger (COCO). ASR is as reported in Table 1. ASR∣Det\mathrm{ASR}_{\vert\mathrm{Det}} (ASR (cond.) in the table) restricts it to the poisoned objects whose original class the backdoored model detects in the paired image without the trigger. Δ=ASR∣Det−ASR\Delta = \mathrm{ASR}_{\vert\mathrm{Det}} - \mathrm{ASR}. Faster denotes Faster R-CNN.

ObjectiveMethodFCOS ASRFCOS ASR (cond.)FCOS Δ\DeltaFaster ASRFaster ASR (cond.)Faster Δ\DeltaDINO ASRDINO ASR (cond.)DINO Δ\Delta
ODABadDet+96.9597.34+0.3998.4697.80-0.6697.6097.07-0.53
Align33.3629.00-4.3638.2333.45-4.7832.1654.98+22.82
Align Rnd55.2451.19-4.0561.9457.73-4.2179.9273.21-6.71
UBA28.6529.16+0.5144.3633.61-10.7597.8997.11-0.78
UBA Box35.1329.54-5.5939.6532.97-6.6897.4396.30-1.13
RMABadDet+99.2899.74+0.4699.4599.67+0.2297.2799.33+2.06
BadDet99.4599.83+0.3899.4899.83+0.3599.2699.71+0.45

Conditioning on untriggered detection changes the ODA-ASR of BadDet+ by at most 1.9 points on MTSD and 0.7 points on COCO, whereas the baselines lose up to 17.0 points on MTSD and 10.8 points on COCO (Align on DINO instead gains 22.8 points, as objects that DINO misses in the clean image are detected once the background triggers are added; it remains the weakest attack on DINO under both measures, and Appendix B.4 now notes this). Under RMA, conditioning raises the ASR of BadDet+ by 0.2-2.9 points.

The strictest criterion, clean-conditioned and class-agnostic, differs from ASR∣Det\mathrm{ASR}_{\vert\mathrm{Det}} only by objects that are detected without the trigger, hidden under the original class, and then detected under another class. Other-class detections accompany at most 2.8% of BadDet+ cases on MTSD and 3.3% on COCO (Table 2.1a in Response 2.1(a)), so even if every one of them fell in this intersection the strict rate would be at most 3 points below ASR∣Det\mathrm{ASR}_{\vert\mathrm{Det}}, which is within a few points of Tables 1 and 2 and far above every baseline except UBA and UBA Box on DINO for COCO, which are already near saturation. The joint occurrence therefore cannot change the ranking or the interpretation. The reported ODA-ASR of BadDet+ reflects objects that are no longer detected, not objects missed before the trigger was added or reassigned to another class, and the ODA gains are not an artefact of the per-class success definition.

Comment 2.2: For PTSD/FCOS RMA, an ASR of 80.59% with a TDR of 39.69% caps the joint success rate at 60.31%, which does not support the claim of "reliable" replacement, so the joint outcome should be reported directly.

Response 2.2: We agree that there is additional nuance present in the interpretation of the results. In the case of PTSD, higher TDR means that separating out each outcome more granularly is useful. As a result, we provide a more detailed breakdown of the joint outcome of RMA, which includes the percentage of frames in which only the target class is detected, only the original class is detected, or both classes are detected. We also report the percentage of frames in which the highest target-class score exceeds the highest original-class score (T>O). We have added this analysis as a new Appendix B.5. Below we provide the results for convenience (Table 2.2), together with a short discussion.

Table 2.2. Joint outcome of RMA on PTSD under random trigger placement (%, IoU 0.50). ASR = Target only + Both and TDR = Both + Original only. T>O is the percentage of frames in which the highest target-class score exceeds the highest original-class score. Bold marks the value that favours the attack. Faster denotes Faster R-CNN.

ModelMethodASRTDRBothTarget onlyOriginal onlyT>O
FCOSBadDet+80.5939.6920.4960.1019.271.58
BadDet73.4884.2559.5913.8924.6639.84
FasterBadDet+85.7728.7715.2270.5513.5579.60
BadDet97.7599.5497.290.462.2539.76
DINOBadDet+80.7819.031.0279.7618.0180.14
BadDet75.2328.694.4270.8124.2773.17
YOLOv5BadDet+67.4334.635.9861.4528.6563.47
BadDet81.2017.881.2979.9116.5980.48

BadDet+ achieves target-class-only detection in 60.1%, 70.5%, 79.8% and 61.5% of frames on FCOS, Faster R-CNN, DINO and YOLOv5. BadDet reaches only 13.9% and 0.5% on FCOS and Faster R-CNN, because most of its ASR comes from frames in which both classes are detected, and its highest target-class score exceeds the original-class score in under 40% of frames (71.6-80.1% for BadDet+). BadDet's high ASR on these two models therefore does not mean the target class dominates the prediction. BadDet+ also leads on DINO, while YOLOv5 remains the exception noted in the paper. Section 5.2 now refers to this analysis, states the target-only rates above directly, and notes that BadDet's higher ASR on PTSD consists mostly of frames in which both classes are detected. It also reports T>O and notes that, where a downstream decision depends on the highest-scoring class, T>O is the rate at which the physical attack changes that decision; target-only frames are the strictest outcome, but the attack affects the decision in 71.6-80.1% of frames on FCOS, Faster R-CNN and DINO. We have also removed "reliable" from the abstract, introduction and contributions where it described BadDet+.

Reviewer_Scd5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper reassesses object-detection backdoor evaluations, highlighting retained original-class detections in region misclassification, limitations of poisoned-set mAP for disappearance attacks, and sensitivity to trigger scaling. It reports that several poisoning-only attacks do not consistently achieve effective attacks with preserved clean performance. Under an outsourced-training threat model, BadDet+ adds an object-matched original-class suppression penalty. Experiments cover four detector families, COCO, MTSD, and physical PTSD images, with penalty ablations and defense evaluations.

Critical strengths
  1. A valuable evaluation diagnosis. Showing that target-class ASR can coexist with retained original-class predictions is important for interpreting RMA results. The qualitative examples, TDR measurements, and separation of instance-level outcomes from dataset mAP make this issue concrete.
  2. An interpretable mechanism with controlled evidence. Table 8 isolates the suppression penalty while holding the poisoned data and training protocol fixed. The large Faster R-CNN TDR reduction and ODA improvements substantiate a useful mechanism. Coverage of multiple detector families and physical images is also valuable; the YOLOv5 exception and defense false-positive rates are openly reported.
Critical weaknesses
  1. W1 — The proposed ODA metric still conflates distinct failure modes. Section 5.1 counts disappearance whenever the original-class detection is absent. An overlapping wrong-class detection can therefore count as successful disappearance even though the object remains detected. The denominator includes all poisonable objects, without stated conditioning on successful detection in the corresponding untriggered image, so ordinary misses can also count as attack successes. For an evaluation-focused paper, class-agnostic disappearance and trigger-induced detection loss must be separated from these alternatives.

  2. W2 — The poisoning-limit conclusion exceeds the evidence. Appendix A defines BadDet's nominal 100% as one poisoned object per eligible image, with eligibility caps—not corruption of every object. Definitions also differ across attacks. These sweeps establish limitations of the tested implementations, not a general detection-versus-classification resistance result. Moreover, Table 10's raw prediction fractions do not establish causal loss dilution: native losses use architecture-specific normalization; for example, Torchvision FCOS normalizes by foreground count. Realized object-level poisoning and effective loss/gradient weighting need to be distinguished, or the mechanism claim narrowed to a hypothesis.

  3. W3 — Threshold robustness results need reconciliation with the main table. Appendix B.1 explicitly claims the Table 2 settings, yet Table 2 reports MTSD/FCOS ODA-ASR of 83.68%, while Figure 6 places it above 90% throughout the plotted threshold range and the text reports at least 91.3% at the lowest threshold. Table 2 also reports FCOS RMA-TDR of 16.96%, whereas the threshold discussion says it never exceeds 10.3%. Different checkpoints, poisoning rates, evaluation cohorts, or thresholds might explain this, but are not identified. These are material to the claimed robustness, not merely formatting issues.

  4. W4 — Relevant training-control prior art is mischaracterized. Section 2.1 describes Lu et al.'s AnywhereDoor as sample-specific and without a deployment strategy. Its original paper instead specifies input-agnostic triggers, full control of detector training, and distribution through model zoos (Sections 1 and 3). Thus training-control attacks already exist in object detection. The contribution needs a more accurate comparison to this setting and a clearer account of what the original-class suppression objective adds. The penalty-off ablation is useful but does not replace this positioning.

Questions for authors
  1. Does ODA-ASR count overlapping wrong-class boxes as successful disappearance? From existing predictions, can you separate class-agnostic disappearance and, for paired digital inputs, disappearance conditional on clean detection?
  2. At nominal 100% poisoning, what fractions of images and ground-truth objects are actually poisoned for each attack after eligibility restrictions? Should the general poisoning-resistance claim be restricted to the evaluated implementations and protocols?
  3. Which checkpoints, poisoning rates, object cohorts, and exact score thresholds underlie Table 2 and Figures 5–6? Please reconcile the FCOS ODA-ASR and RMA-TDR discrepancies.
  4. How does BadDet+ differ from AnywhereDoor's existing training-control threat model, and what supports the sample-specific characterization in Section 2.1?
  5. Are Table 10's losses measured before or after each native loss's normalization? What links the raw prediction counts to effective gradient contribution? If that link is not established, can loss dilution remain an explanatory hypothesis rather than an established cause?
6 comments

We thank the reviewer for their constructive and detailed comments, which have helped us to improve the paper. Below we respond to each comment in turn. Where several reviewers raised the same point, we repeat the full response in each thread so that each can be read independently. We are keen to engage in discussion as much as possible during the rebuttal period to ensure a positive outcome, and we are happy to address any residual concerns or provide further clarifications.

Comment 1.1(a): (W1, Q1) ODA-ASR counts an overlapping other-class box as a successful disappearance, so class-agnostic disappearance should be separated from misclassification.

Response 1.1(a): You are correct. ODA-ASR counts an object as hidden when no original-class box overlaps it, so an overlapping box of another class would still count as a success. RMA evaluation has a similar issue, since TDR only checks whether the original class is detected, and ASR only checks whether the target class is detected.

To quantify how often this occurs, we added the other-class detection rate (OCD) to a new Appendix B.4 (reproduced below as Table 1.1a): the percentage of targeted objects with a prediction (IoU ≥0.5\geq 0.5) of a class other than the original (and, for RMA, the target), for all methods and detectors on MTSD, PTSD and COCO.

Table 1.1a. Other-class detection (OCD) under random trigger placement on MTSD and PTSD and centre placement on COCO: the percentage of cases in which a prediction overlaps the ground truth (IoU ≥0.5\geq 0.5) and has a class other than the original and, for RMA, the target. Bold marks the lowest value in each column. Faster denotes Faster R-CNN. A dash marks configurations not evaluated in Tables 1 and 2.

ObjectiveMethodFCOS MTSDFCOS PTSDFCOS COCOFaster MTSDFaster PTSDFaster COCODINO MTSDDINO PTSDDINO COCOYOLOv5 MTSDYOLOv5 PTSD
ODABadDet+0.520.533.310.590.800.130.871.030.182.803.86
Morph2.428.30-1.901.48-1.030.27-4.280.68
Align--0.72--0.75--0.41--
Align Rnd--0.24--0.25--0.29--
UBA1.822.212.381.470.152.882.463.390.141.437.24
UBA Box2.061.982.021.740.191.340.630.000.131.357.01
RMABadDet+0.260.463.271.114.910.410.000.000.170.211.79
BadDet3.122.2114.0812.418.725.380.000.000.270.040.19
Morph10.3512.98-9.286.93-0.000.00-0.260.08

Other-class boxes are rare: under ODA they accompany at most 1.03% of BadDet+ cases on FCOS, Faster R-CNN and DINO on MTSD and PTSD, 3.86% on YOLOv5, and at most 3.31% on COCO, versus up to 8.3% for the baselines. A class-agnostic success criterion could therefore lower the reported ODA-ASR of BadDet+ by at most these amounts, which does not change the ranking in Tables 1 and 2 or its interpretation. Under RMA, BadDet+ leaves a third class on at most 1.11% (MTSD), 4.91% (PTSD) and 3.27% (COCO) of cases, versus up to 14.08% for the baselines, which agrees with the lower TDR we report for BadDet+. The reported ODA-ASR of BadDet+ therefore reflects objects that are no longer detected, not objects reassigned to another class. Response 1.1(b) combines this with conditioning on detection without the trigger.

Comment 1.1(b): (W1, Q1) The ODA-ASR denominator includes all poisonable objects without conditioning on detection in the untriggered image, so ordinary misses can count as attack successes.

Response 1.1(b): You are correct. The denominator includes all poisonable objects, so an object that is already missed without the trigger counts as a success. To quantify how often this occurs, we added ASR∣Det\mathrm{ASR}_{\vert\mathrm{Det}} to the new Appendix B.4 (reproduced below as Tables 1.1b and 1.1c): ASR restricted to the objects whose original class the backdoored model detects in the paired untriggered image (IoU ≥0.5\geq 0.5, default score threshold). PTSD provides no paired untriggered images, so ASR∣Det\mathrm{ASR}_{\vert\mathrm{Det}} is computed on MTSD and COCO. It is also reported for RMA for completeness.

Table 1.1b. ASR conditioned on detection without the trigger (MTSD, random trigger placement). ASR is as reported in Table 2. ASR∣Det\mathrm{ASR}_{\vert\mathrm{Det}} (ASR (cond.) in the table) restricts it to the poisoned objects whose original class the backdoored model detects in the paired image without the trigger. Δ=ASR∣Det−ASR\Delta = \mathrm{ASR}_{\vert\mathrm{Det}} - \mathrm{ASR}. Faster denotes Faster R-CNN.

ObjectiveMethodFCOS ASRFCOS ASR (cond.)FCOS Δ\DeltaFaster ASRFaster ASR (cond.)Faster Δ\DeltaDINO ASRDINO ASR (cond.)DINO Δ\DeltaYOLOv5 ASRYOLOv5 ASR (cond.)YOLOv5 Δ\Delta
ODABadDet+83.6881.78-1.9089.3889.22-0.1692.3191.85-0.4687.0885.72-1.36
Morph7.444.53-2.914.212.30-1.9157.4440.43-17.0149.5143.48-6.03
UBA32.7925.03-7.760.000.000.008.081.34-6.7422.6313.47-9.16
UBA Box27.5121.35-6.163.930.82-3.1187.2283.51-3.7117.327.62-9.70
RMABadDet+93.1396.01+2.8897.0497.99+0.9590.4393.28+2.8587.0489.15+2.11
BadDet84.9084.65-0.2593.9699.06+5.1083.3986.64+3.2593.2593.39+0.14
Morph36.9434.40-2.5438.4138.60+0.1914.0314.15+0.1258.6158.77+0.16

Table 1.1c. ASR conditioned on detection without the trigger (COCO). ASR is as reported in Table 1. ASR∣Det\mathrm{ASR}_{\vert\mathrm{Det}} (ASR (cond.) in the table) restricts it to the poisoned objects whose original class the backdoored model detects in the paired image without the trigger. Δ=ASR∣Det−ASR\Delta = \mathrm{ASR}_{\vert\mathrm{Det}} - \mathrm{ASR}. Faster denotes Faster R-CNN.

ObjectiveMethodFCOS ASRFCOS ASR (cond.)FCOS Δ\DeltaFaster ASRFaster ASR (cond.)Faster Δ\DeltaDINO ASRDINO ASR (cond.)DINO Δ\Delta
ODABadDet+96.9597.34+0.3998.4697.80-0.6697.6097.07-0.53
Align33.3629.00-4.3638.2333.45-4.7832.1654.98+22.82
Align Rnd55.2451.19-4.0561.9457.73-4.2179.9273.21-6.71
UBA28.6529.16+0.5144.3633.61-10.7597.8997.11-0.78
UBA Box35.1329.54-5.5939.6532.97-6.6897.4396.30-1.13
RMABadDet+99.2899.74+0.4699.4599.67+0.2297.2799.33+2.06
BadDet99.4599.83+0.3899.4899.83+0.3599.2699.71+0.45

Conditioning on untriggered detection changes the ODA-ASR of BadDet+ by at most 1.9 points on MTSD and 0.7 points on COCO, whereas the baselines lose up to 17.0 points on MTSD and 10.8 points on COCO (Align on DINO instead gains 22.8 points, as objects that DINO misses in the clean image are detected once the background triggers are added; it remains the weakest attack on DINO under both measures, and Appendix B.4 now notes this). Under RMA, conditioning raises the ASR of BadDet+ by 0.2-2.9 points.

The strictest criterion, clean-conditioned and class-agnostic, differs from ASR∣Det\mathrm{ASR}_{\vert\mathrm{Det}} only by objects that are detected without the trigger, hidden under the original class, and then detected under another class. Other-class detections accompany at most 2.8% of BadDet+ cases on MTSD and 3.3% on COCO (Table 1.1a in Response 1.1(a)), so even if every one of them fell in this intersection the strict rate would be at most 3 points below ASR∣Det\mathrm{ASR}_{\vert\mathrm{Det}}, which is within a few points of Tables 1 and 2 and far above every baseline except UBA and UBA Box on DINO for COCO, which are already near saturation. The joint occurrence therefore cannot change the ranking or the interpretation. The reported ODA-ASR of BadDet+ reflects objects that are no longer detected, not objects missed before the trigger was added or reassigned to another class, and the ODA gains are not an artefact of the per-class success definition.

Comment 1.2: (W2a, Q2) Nominal 100% poisoning does not poison every object, so the effective image- and object-level poisoning fractions should be reported and the poisoning-limit conclusion restricted to the evaluated implementations and protocols.

Response 1.2: We agree with both points. For BadDet, a nominal rate of 100% means one poisoned object in every eligible image, so not every object is poisoned. To test whether this definition explains BadDet's limits, we reran BadDet with the methodology used by UBA, which poisons multiple objects per image, so that the ratio is defined over all annotated objects. We swept the poisoning ratio from 10% to 50%. We have added this sweep to Appendix C (Fig. 8c). Below we provide the relevant results (Table 1.2a) as well as a short discussion for your convenience.

Table 1.2a. BadDet RMA on MTSD (random placement) with the poisoning ratio defined over images (one poisoned object per image, as in the paper) or over objects. Clean baseline mAP: FCOS 58.5, Faster R-CNN 55.3, DINO 59.3. Bold marks rows in which BadDet improves on BadDet+ in both ASR and TDR. Faster denotes Faster R-CNN.

RatioDefined overFCOS mAPFCOS ASRFCOS TDRFaster mAPFaster ASRFaster TDRDINO mAPDINO ASRDINO TDR
10%Images55.0152.2292.1849.6169.8097.3456.0661.6337.48
Objects54.4278.8771.7349.2292.0997.1855.8685.8414.93
20%Images53.7471.7684.5548.6689.5297.5156.0882.5917.94
Objects52.5889.3949.1045.1196.7196.4956.1292.516.07
30%Images53.5384.9066.9647.4893.9697.8754.1083.3914.35
Objects49.7293.4640.8041.3197.5695.3452.7591.875.47
40%Images51.9686.5060.0446.3196.4596.4554.4090.5912.08
Objects42.9894.7034.6937.0997.8689.1848.3696.452.87
50%Images52.9092.0148.4944.7296.8096.6354.8191.308.53
Objects25.7495.8116.1724.7598.1658.8523.5293.333.42
BadDet+-55.8693.1316.9653.4697.049.1353.3590.435.39

Poisoning more objects does strengthen BadDet, but at a cost in clean performance. Reaching the TDR of BadDet+ on FCOS requires a 50% object-level ratio, at which mAP collapses to 25.7 (BadDet+: 55.9). On Faster R-CNN, TDR never falls below 58.9%, even at 24.8 mAP. DINO is the other detector on which poisoning alone suffices: object-level BadDet matches BadDet+ at a 20% ratio (92.5% ASR, 6.1% TDR and 56.1 mAP against 90.4%, 5.4% and 53.4) and exceeds it on both ASR and TDR at 40%, at a cost of 5 mAP. Section 5.2 and Appendix C now state this.

These results sharpen our conclusions, and we have revised the wording to match. We agree that "more significant threat" read as a claim about likelihood. The revision now qualifies it: outsourced training is the more serious threat for object detection in the sense that matters for defence, because it yields a backdoor that is consistent across detector architectures, datasets and objectives, whereas the evaluated poisoning-only attacks succeed only on particular detectors, and we argue that detection defences should be assessed under this threat model for that reason (Sections 5.2 and 6). We have also replaced "maximal poisoning" and "to its limit" in Sections 1 and 3, removed the "100%" framing from the abstract and introduction, and added the effective-rate qualifier to the sweep in Section 3; our claim concerns the evaluated poisoning-only attacks and protocols, and that scope is now explicit throughout (Sections 1, 3, 5.2 and 6, and Appendix C). What remains is a claim about consistency. Across the 38 detector-dataset-objective settings in Tables 1, 2 and 13 (22 trained models, with the PTSD rows re-evaluating the MTSD models), BadDet+ leads the best poisoning-only attack by at least 5 points in 30: 17 of 19 for ODA-ASR and 13 of 19 for RMA-TDR. Where it does not, poisoning alone is competitive: on DINO for both objectives and on YOLOv5 for RMA, and Sections 1 and 5.2 now say so. Poisoning-only success is therefore real but conditional on the detector: BadDet succeeds only on DINO and YOLOv5, UBA Box only on DINO, and no baseline succeeds on FCOS or Faster R-CNN for either objective without sacrificing clean accuracy, whereas in image classification poisoning-only attacks succeed consistently across architectures. BadDet+ succeeds on all four detectors for both objectives, and Section 5.3 shows that none of the evaluated defences stops it.

Appendix C (Table 11) now reports the effective poisoning rates on MTSD. At nominal 100%, 81.1% of training images are poisoned for BadDet and BadDet+ RMA (one object in each), 86.2% for BadDet+ ODA, and 80.1% of annotated objects for UBA. Image-level and object-level rates have different denominators, so each attack is reported under its native definition, and Section 3 now states the image-level rate; the object-level sweep above is the comparable experiment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.