.png){kind=icon}
.png){kind=icon}
Return
CaDe: Adaptive Sparse Causal Decoupling for Adversarially Robust Object Detection via Hierarchical Stability Constraints
.png){kind=icon}
PRE
AI
W
Y
H
P
H
DOI：10.1109/tifs.2026.3715119.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Object detectors are widely deployed in safety-critical systems, but their robustness is significantly threatened by physical adversarial patch (AP) attacks. While various defense mechanisms have been proposed in recent years to mitigate such attacks, existing methods primarily focus on suppressing perturbation-induced degradation in the feature space, without systematically analyzing how these perturbations propagate across the hierarchical structure of deep neural networks (DNNs) and disrupt the decision-making process. This paper examines the primary destructive mechanism of APs, which largely involves the disruption of operational causal consistency within the model rather than directly corrupting feature values. Specifically, perturbations disrupt the operational causal consistency of features through layer-wise propagation, disrupting the model’s decision pipeline and inducing cross-layer coupling interference, ultimately leading to reduced robustness. To target this mechanism, we propose a defense method for robust object detection called CaDe. CaDe mitigates the hierarchical propagation of residual perturbations through hierarchical stability constraints and adaptive sparse causal decoupling strategies, fundamentally enhancing the model’s robustness. Theoretical analysis demonstrates that CaDe possesses global exponential stability, preserving task-relevant semantic representations while sparsely separating adversarial components. Experimental results show that CaDe can effectively defend against Hiding Attack (HA) and Appearing Attack (AA), achieving improvements of 6.19% and 5.53% in mean Average Precision (mAP) compared with the best baseline method, respectively. Additionally, the mAP on benign samples only decreases by 0.14% compared with the baseline. Furthermore, CaDe incurs minimal additional computational overhead, enabling significant improvements in robustness while maintaining real-time performance.
Keywords:
Object detection
adversarial patch defense
operational causal decoupling
hierarchical prediction
Journal
IF:
8
Papers:
5.2K
Citations:
2.3W
Organization
Cited Papers