ConvNet-Generated Adversarial Perturbations for Evaluating 3D Object Detection Robustness
Bibliographic record
Abstract
This paper presents a novel adversarial Convolutional Neural Network (ConvNet) method for generating adversarial perturbations in 3D point clouds, enabling gradient-free robustness evaluation of object detection systems at inference time. Unlike existing iterative gradient methods, our approach embeds the ConvNet directly into the detection pipeline at the voxel feature level. The ConvNet is trained to maximize detection loss while maintaining perturbations within sensor error bounds through multi-component loss constraints (intensity, bias, and imbalance terms). Evaluation on a Sparsely Embedded Convolutional Detection (SECOND) detector with the KITTI dataset shows 8% overall mean Average Precision (mAP) degradation, while CenterPoint on NuScenes exhibits 24% weighted mAP reduction across 10 object classes. Analysis reveals an inverse relationship between object size and adversarial vulnerability: smaller objects (pedestrians: 13%, cyclists: 14%) show higher vulnerability compared to larger vehicles (cars: 0.2%) on KITTI, with similar patterns on NuScenes, where barriers (68%) and pedestrians (32%) are most affected. Despite perturbations remaining within typical sensor error margins (mean L2 norm of 0.09% for KITTI, 0.05% for NuScenes, corresponding to 0.9-2.6 cm at typical urban distances), substantial detection failures occur. The key novelty is training a ConvNet to learn effective adversarial perturbations during a one-time training phase and then using the trained network for gradient-free robustness evaluation during inference, requiring only a forward pass through the ConvNet (1.2-2.0 ms overhead) instead of iterative gradient computation, making continuous vulnerability monitoring practical for autonomous driving safety assessment.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.002 | 0.007 |
| Meta-epidemiology (narrow) | 0.002 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.002 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".