面向场景文本检测模型的迁移对抗攻击

Transfer adversarial attacks on scene text detection models

针对场景文本检测算法的攻击算法不能有效兼顾迁移性、隐蔽性和攻击效果的问题,提出MIFGSM-W攻击算法。算法提出通用概率图,引入动量项获取稳定的梯度更新方向;使用可微函数替代标准二值化函数,构造损失函数;引入变量,并提出改进的优化策略,约束扰动;提出个体攻击算法及通用攻击算法。在多个数据集上实验,结果表明:该攻击算法能够有效攻击EAST,Textbox++,Craft,DBNet场景文本检测模型,且生成的对抗样本兼顾迁移性和视觉隐蔽性。

To address the issues of transferability,stealthiness,and attack effectiveness in attacking scene text detection algorithms,we proposed a MIFGSM-W attack algorithm.The MIFCSM-W attack algorithm introduced a general probability map and a momentum term to obtain stable gradient update directions.Differentiable functions were used instead of standard binarization functions to construct the loss function.Variable was introduced,and an optimization strategy was proposed to improve variable and constrain the perturba-tion.Individual attack algorithms and a general attack algorithm were proposed.The results of our experiments on multiple datasets demonstrate that the proposed attack algorithm successfully targets scene text detection models such as EAST,Textbox++,Craft,and DBNet.Moreover,the generated adversarial samples exhibit both transferability and visual stealthiness.