| 摘要: |
| 对抗训练作为提升深度神经网络对抗鲁棒性的核心策略, 在图像分类任务中已得到广泛关注, 但在目标检测领域中的研究较为匮乏. 传统对抗训练通常依赖投影梯度下降法(projected gradient descent, PGD)开展模型的鲁棒优化, 然而对抗样本的迭代大幅延长了模型训练周期, 成为限制对抗训练在目标检测这类计算密集型任务中实际部署的主要瓶颈. 针对这个问题, 提出一种基于Nesterov加速梯度 (Nesterov’s accelerated gradient, NAG)的对抗训练方法, 通过引入NAG动量机制加速算法收敛, 该方法在得到与PGD所训练模型精度相当的同时, 显著加快了对抗训练速度. 此外, 目标检测与图像分类最主要的区别在于目标边界框定位. 然而观察到现有方法仍侧重于学习基于分类损失产生的对抗样本, 忽视了定位在目标检测中的特殊性. 设计一种自适应损失重加权策略, 以均衡训练中不同任务所衍生对抗样本的数量占比, 促进模型聚焦定位以增强鲁棒性. 在PASCAL VOC和MS-COCO两个公开目标检测数据集上与现有的先进目标检测对抗训练方法进行实验, 对比验证了所提方法的有效性. |
| 关键词: 目标检测 对抗训练 动量优化 多任务损失 |
| DOI：10.13328/j.cnki.jos.007528 |
| 分类号:TP18 |
| 基金项目:国家自然科学基金(62576351, 62076252, 62106281); 中国博士后科学基金 (2024M764294); 湖南省研究生科研创新项目(CX20240105) |
|
| Adversarial Training Method for Object Detection Based on Momentum Acceleration and Task Balancing |
|
LONG Sheng1, LIN Chen2, TAO Wei1,3, ZHANG Jun1, TAO Qing4
|
|
1.Laboratory for Big Data and Decision (National University of Defense Technology), Changsha 410073, China;2.Joint Laboratory of Data Science and Business Intelligence (Southwestern University of Finance and Economics), Chengdu 610074, China;3.Center for Strategic Assessment and Consulting, Academy of Military Sciences, Beijing 100091, China;4.Department of Information Engineering, Army Academy of Artillery and Air Defense Forces, Hefei 230031, China
|
| Abstract: |
| Adversarial training, a key strategy for enhancing the adversarial robustness of deep neural network (DNNs), has been widely studied in image classification but lacks sufficient research in object detection. Traditional adversarial training often relies on projected gradient descent (PGD) for robust optimization of models. However, the iterative process of generating adversarial examples greatly prolongs model training, becoming a major bottleneck for deploying adversarial training in computationally intensive tasks like object detection. To address this, this study proposes an adversarial training method based on Nesterov’s accelerated gradient (NAG). By introducing the NAG momentum mechanism, algorithm convergence is accelerated. This method maintains detection accuracy comparable to PGD-trained models while significantly improving adversarial training efficiency. In addition, the main difference between object detection and image classification lies in object bounding box localization. However, it is observed that existing methods still focus on learning adversarial examples generated from classification loss, while neglecting the particularity of localization in object detection. To address this, an adaptive loss re-weighting strategy is designed to balance the number of adversarial examples derived from different tasks during training, thus enabling the model to focus on localization to enhance robustness. Experiments on the PASCAL VOC and MS-COCO datasets demonstrate the effectiveness of the proposed method compared with existing advanced adversarial training approaches for object detection. |
| Key words: object detection adversarial training momentum optimization multi-task loss |
