- Rating
- 3
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
The paper proposes to fix the malicious input in order to leverage the information instead of discarding it. The proposed method (Gecko) first determines where corruption occurs and how the affected information should be restored in the feature space. Then it verifies the repaired features in the prediction space to identify any remaining harmful effects. Results show SOTA reduction in ASR.
Critical strengths
The paper follows a strong scientific methodology. I appreciate the thoroughness of the experiments. We have the full analysis up to the latency overhead, which is nice!
The paper proposes a novel solution that repairs instead of discarding and performs great.
Critical weaknesses
I am rated the paper as weak accept because the beginning is a bit rough to follow. I think it would benefit of being clearer. See my comments below:
- Section 3.1 (system model): you need to explain how are information exchanged between CAVs. You mention "after feature alignment and before collaborative fusion" but a reader won't necessarily know what these are.
- Section 3.2 (threat model): please use the conventional threat model used in CAV security literature. You assume multiple non-colluding attackers, who perturb their intermediate feature map. The perturbation(s) must be explained in detail (you only said "suppressing safety-critical objects" "perception errors"). This is not sufficient to understand what they perturb (how, how much, always/periodically, etc.). Later in Section 4.3, you mentioned "delta is sampled from a diverse set of training-time perturbations": what are these perturbations? In section 5.1, you mention PGD, BIM, MOR, TOR and the defenses (ROBOSAC, LUCIA, etc.) without explaining the attacks (especially MOR and TOR), or the differences between the defense. You need a short related work section.
- Identification of corrupted cells: can you only do it if the ego vehicle receives information from multiple vehicles? What is the requirement in terms of number of contributors/sources for Gecko to work?
- (Section 4.3) Detail what is "task-relevant information" (used in L_task), and how is it selected.
- You pretrained a VAE to provide a recovery reference, so how sensitive to the training data is it? You must have a good generalization (or enough training data) in order to have a reliable F_vae, no?
- During the LOOC consensus, do you recursively leave one out (if Epsilon = 5, the consensus is from 4 CAVs) or do you cumulatively remove one by one (start with 5, then 4, 3, 2, 1? I am unclear about this process.
- From the short paragraph on "feature repair" in section 4.3, it is unclear how it works. After you locate the corrupted cell in the map, you simply replace it by F_vae? Or do you try purification model on Fi(x)? How can you know the correction to apply to the feature to eliminate it? Or you do not repair, you simply ignore that contribution from CAVi to the corrupted cell, but use CAVi contributions to all the other cells? Actually, when reading Section 4.4., I wonder if what I thought was done in 4.3 (localization of corrupted cell and CAV) isn't the same as "prediction-space verification". It seems you need to explain better each steps and if 4.4 is really needed.
- In terms of robustness against more attackers, what is the ratio attacker/benign? If it is max 4 out of 20 CAV, that is normal. But if it is 4 out of 7 CAV (majority of malicious CAV), then please explain.
- Section 5.4 (Ablation): Should I understand the Table 3, as progressively removing each component, or each row represents remove one component (but the three other remains)? As you concluded, "anomaly" is the most important component, while "deviation" has a slightly higher benefit that "suppression" and "correction". Do you also have the latency results if we only do "anomaly" and none of the other components?
Minor comments:
- Figure 5: the choice of the graphics (fill in) is not ideal. I can't figure out which is which.
- Are you planning to release the trained VAE (as it seems essential)?
- I know we are not supposed to review Appendix, but Figure 7 is very useful. Any option to pull it in the main body?
- I like the latency analysis. I would put that 100ms overhead in perspective to the latency of the full perception pipeline and maximum acceptable latency. 100ms might be acceptable.
Questions for authors
explain the weaknesses listed earlier.
8 comments
Response to Reviewer Riqu (Part 2/4)
Q3. Identification of corrupted cells and minimum collaboration requirements.
Thank you for raising this concern. Gecko does not require multiple collaborating CAVs to identify and repair corrupted features. It can operate with one ego CAV and only one collaborating CAV. The results in Table 5 of the original manuscript also demonstrate that effective repair of compromised features can still be achieved with only one CAV.
Specifically, when a sufficient number of other collaborating CAVs can provide reliable collaborative references, Gecko combines VAE anomaly evidence, cross-vehicle deviation evidence, and suppression evidence to localize corrupted regions, and uses collaborative consensus for feature repair. When the number of collaborating CAVs is limited and reliable collaborative references are unavailable, Gecko relies on VAE anomaly evidence and suppression evidence to localize anomalies, and uses VAE reconstructions as a reference for feature repair when the corresponding anomaly conditions are satisfied.
Neither VAE anomaly evidence nor suppression evidence requires multiple collaborating CAVs, although they are computed differently. VAE anomaly evidence is obtained by comparing the received features with the VAE-reconstructed features. Since the VAE reconstruction is generated using the features of the CAV under inspection and the ego CAV, its computation does not require other collaborating CAVs to observe the same spatial regions. Suppression evidence, on the other hand, requires a collaborative reference. However, when the system contains only one collaborating CAV, the ego CAV alone provides this reference after excluding the CAV under inspection. Therefore, suppression evidence can still be computed without requiring a third CAV.
Q4. Detail what is "task-relevant information" (used in L_task), and how is it selected.
We thank the reviewer for this question. The task-relevant information in refers to the object classification and bounding-box regression responses of the collaborative perception model, which capture the information needed for object classification and localization. Rather than selecting specific objects, anchors, or spatial regions, we use the complete raw outputs of both detection branches. Specifically, during VAE training, we feed the clean features and their VAE-reconstructed counterparts into the same frozen collaborative perception model and minimize the mean absolute error between their classification and regression outputs. This encourages the reconstructed features to preserve the detection behavior of the original clean features.
Q5. Generalization and training data sensitivity of VAE.
Thank you very much for your concern regarding the data sensitivity of the VAE. To address your concern, we evaluated the performance of our method under different amounts of training data. The experimental results are shown in Table R1.
As shown in Table R1, Gecko maintains relatively stable performance across different VAE training data sizes. Under MOR and TOR attacks, [email protected] remains at 0.90, while [email protected] varies only between 0.81 and 0.82. MOR ORR and TOR ASR also remain within narrow ranges of 2.32%–2.48% and 2.58%–2.78%, respectively. Under PGD attacks, although ORR increases to 4.62% with the full training set, [email protected] and [email protected] remain comparable at 0.88 and 0.80. Notably, using only 10% of the training data achieves performance comparable to using the full set, indicating that Gecko is not highly sensitive to VAE training data size.
Table R1: Sensitivity of Gecko to VAE training data size on the OPV2V test set. AP is reported on a 0–1 scale, while ORR and ASR are reported as percentages. Higher AP and lower ORR/ASR indicate better defense performance.
| VAE training data | MOR ORR ↓ | MOR [email protected] ↑ | MOR [email protected] ↑ | TOR ASR ↓ | TOR [email protected] ↑ | TOR [email protected] ↑ | PGD ORR ↓ | PGD [email protected] ↑ | PGD [email protected] ↑ |
|---|---|---|---|---|---|---|---|---|---|
| 10% | 2.48% | 0.90 | 0.81 | 2.58% | 0.90 | 0.82 | 2.32% | 0.89 | 0.80 |
| 25% | 2.32% | 0.90 | 0.81 | 2.78% | 0.90 | 0.81 | 2.20% | 0.89 | 0.81 |
| 50% | 2.37% | 0.90 | 0.82 | 2.58% | 0.90 | 0.82 | 2.22% | 0.90 | 0.82 |
| 100% | 2.34% | 0.90 | 0.82 | 2.58% | 0.90 | 0.82 | 4.62% | 0.88 | 0.80 |
We have also added this sensitivity analysis to the revised manuscript to further validate the generalization capability and robustness of Gecko.
Response to Reviewer Riqu (Part 3/4)
Q6. Clarification of leave-one-out consensus construction.
We thank the reviewer for pointing out this lack of clarity. We adopt an independent leave-one-out consensus construction strategy, in which only the CAV currently under inspection is excluded each time, without cumulatively excluding CAVs that have been inspected previously.
Specifically, when inspecting CAV , Gecko excludes only the current CAV and constructs the corresponding collaborative consensus using the ego CAV and other eligible collaborating CAVs. The same procedure is independently repeated for each CAV under inspection. For example, when the system consists of five CAVs in total (including the ego CAV), the consensus for inspecting any non-ego CAV is constructed using the remaining four CAVs. The consensus constructions for different CAVs under inspection are independent of each other; there is no cumulative removal process in which the number of CAVs is progressively reduced from five to four, three, two, and finally one. We have further clarified this consensus construction process in Section 4.3.
Q7. Clarification of feature repair and its relationship to prediction-space verification.
Thank you very much for raising this concern. To address these questions, we clarify three aspects: feature repair mechanism, repair source and strength, and the necessity of prediction-space verification.
Feature Repair Mechanism. Gecko selectively repairs corrupted features based on the anomaly level and cross-vehicle observation support at each spatial location, rather than directly replacing all corrupted features with VAE reconstructions or simply discarding the contributions of malicious CAVs in corrupted regions. Specifically, Gecko first combines VAE anomaly evidence, cross-vehicle deviation evidence, and suppression evidence to compute a corruption score for each feature cell. It then uses this score to determine the repair strength and selects an appropriate repair source. In this way, Gecko can repair corrupted regions while preserving as much useful information as possible from the original features that has not been affected by attacks.
Repair Source and Strength. For the selection of the repair reference, Gecko adopts different strategies according to the cross-vehicle spatial support at each location. When other collaborative CAVs can provide effective spatial support, the collaborative consensus constructed by the ego vehicle and other qualified collaborative CAVs (after excluding the CAV under inspection) is used as the repair reference. When effective cross-vehicle spatial support is lacking, the VAE-reconstructed features are used as the repair reference if the VAE anomaly score exceeds a predefined threshold; otherwise, the original features are retained.
After determining the repair reference, Gecko further controls the repair strength according to the contamination score , rather than performing a complete replacement. The specific computation is as follows: where denotes the original features, is the selected repair reference, and represents the contamination degree at that location. A higher contamination score makes the repaired features closer to the selected repair reference, while a lower score preserves more of the original features. In other words, Gecko adaptively restores the contaminated features by leveraging the references provided by VAE reconstruction and collaborative consensus.
Necessity of Prediction-Space Verification. Prediction-Space Verification mainly has two roles. 1) it detects and repairs corruption missed by feature-space localization, particularly from stealthy attacks that leave weak feature-space evidence but significantly affect detection results. 2) It evaluates whether feature-space repair produces reliable downstream predictions and further corrects any residual abnormal effects. To further evaluate the role of prediction-space verification, we compared Gecko with and without the Prediction-Space Verification component. As shown in Table R2, under highly stealthy adaptive attacks, feature-space localization and repair alone cannot effectively recover from the attacks. Only through prediction-space verification can such stealthy attacks be detected by examining the detector responses. This is because some stealthy attacks do not exhibit sufficiently significant anomalies in the feature space. Furthermore, prediction-space verification is triggered in 35.30% of the evaluated frames, indicating its ability to identify feature maps that require further repair.
Table R2: Ablation of prediction-space verification under adaptive PGD attacks on OPV2V.
| Defense Configuration | [email protected]↑ | [email protected]↑ | Triggered Frames |
|---|---|---|---|
| Gecko w/o Prediction-Space Verification | 0.55 | 0.49 | — |
| Gecko w/ Prediction-Space Verification | 0.79 | 0.69 | 724 (35.30%) |
Response to Reviewer Riqu (Part 1/4)
Q1. System model.
We thank the reviewer for this valuable suggestion. In an intermediate-fusion collaborative perception system, each CAV extracts intermediate feature maps from its local sensor observations, and collaborating CAVs transmit these features to the ego CAV. Upon receiving them, the ego CAV uses the relative poses between vehicles to align the feature maps into its own coordinate system, ensuring spatial consistency across different viewpoints. The aligned features are then fused with the ego feature through a collaborative fusion module, and the resulting representation is used for object detection.
Under this system model, Gecko is deployed immediately before the intermediate fusion stage to defend against potential attacks from other collaborating CAVs. As an independent defense module, Gecko analyzes each feature map received by the ego CAV from other CAVs, identifies and repairs spatial regions that may be affected by adversarial attacks, and then passes the repaired features to the original fusion algorithm. Therefore, Gecko does not require any modifications to the original collaborative perception system's fusion architecture or object detection module. Additionally, we have revised Section 3.1 of the manuscript.
Q2. Threat model.
We thank the reviewer for pointing out this issue. To address these concerns, we will provide further clarification on four aspects: attacker capabilities, attack objectives and execution, training-time perturbations, and attack configurations.
Attacker Capabilities. We follow the threat model widely adopted in prior adversarial collaborative perception studies, assuming that an attacker can control one or more legitimate non-ego CAVs and manipulate the intermediate features they transmit to the ego CAV over the network. Similar to prior studies, including Tu et al. (ICCV 2021) and LUCIA (USENIX Security 2025), we assume that attackers have access to the collaborative perception model used by the CAVs and can generate adversarial perturbations through gradient-based optimization. In addition, we consider both single-attacker and multi-attacker scenarios, where multiple attackers can independently generate perturbations.
Attack Objectives and Execution. We consider two types of attack objectives: untargeted attacks that aim to degrade overall object detection performance, and object removal attacks that aim to make one or more specific objects disappear from the final perception results. Regarding attack timing, we adopt a continuous attack setting, where attackers continuously manipulate the intermediate features of compromised CAVs throughout the entire evaluation sequence, from the first frame to the last frame. This setting allows us to evaluate the effectiveness of our defense under a worst-case attack scenario.
Training-Time Perturbations. The δi in Section 4.3 refers to synthetic feature perturbations used for VAE training, which differ from the specific attacks used to evaluate Gecko during testing. To improve the generalization ability of the VAE repair model in Gecko against different unseen perturbations, we adopt diverse feature corruption strategies to simulate attacks during offline training. Specifically, these strategies include random feature noise and gradient-based perturbations with varying strengths, where the perturbation magnitude, spatial structure, and optimization objectives are randomly varied.
Attack and Defense Configurations. We consider four attacks: PGD, BIM, MOR, and TOR. PGD and BIM iteratively optimize adversarial perturbations to degrade detection performance, with PGD using random initialization and an L∞ constraint (ϵ=0.3), while BIM starts from zero. Both use 10 iterations with a step size of 0.1. MOR (Mass Object Removal) and TOR (Targeted Object Removal) aim to suppress all detected objects and a specific target object, respectively. For defense comparisons, ROBOSAC uses sampling-based consensus verification, while CP-Guard employs recursive consistency checking to identify malicious CAVs. LUCIA estimates feature-consistency-based trust scores and uses them to adjust fusion attention. PRBI combines temporal consistency with Bayesian inference, whereas Cerberus employs multi-dimensional feature verification.
We have also revised Sections 3.2, 4.3, and 5.1 of the manuscript.