Boosting deep detector efficiency and robustness through detection discriminant reorganization and compression

Deep neural networks have transformed numerous fields, yet their deployment in both resource-constrained and safety-critical settings remains hindered by large model sizes and vulnerability to adversarial attacks. Few studies address both efficiency and adversarial robustness simultaneously, with most focusing on classification tasks, leaving the complex challenges of detection tasks unaddressed. To tackle these challenges, we propose a proactive deep detection discriminant analysis (D3A) framework that boosts both the inference efficiency (e.g., model compactness and latency) and adversarial robustness of deep visual detectors while preserving their performance. Unlike after-the-fact compression methods, our training-time approach actively extracts the detection utility, isolates and condenses it into fewer dimensions aligned with latent filters to enable structured pruning. The process removes easily attackable, interfering dimensions and simultaneously mitigates the long-standing misalignment between the classification and localization subtasks, thereby supporting more effective adversarial training of detectors. We demonstrate the efficacy of our approach across a variety of modern visual detectors and benchmark it against multiple state-of-the-art compression methods on the KITTI and COCO datasets. For example, on KITTI, our compressed YOLOX-S detector, with 20% fewer parameters, outperforms the original by 1.9% mAP on clean images and shows robustness improvements of up to 11.4% and 24.8% under standard and adversarial training, respectively.