首页|一种随机束搜索文本攻击黑盒算法

一种随机束搜索文本攻击黑盒算法

扫码查看
针对现有对抗样本生成算法容易陷入局部最优解的问题,提出了一种名为R-attack的算法,通过束搜索和随机元来提高攻击成功率.利用束搜索在同义词空间中寻找最优解,增加对抗样本的多样性,进而提高攻击的效率,同时,在迭代搜索过程中引入随机元素,避免过早陷入局部最优解,从而有效提高攻击成功率.在3个数据集上对2个模型进行了对抗攻击实验,实验结果表明,使用R-attack算法能够有效提高对抗样本的攻击成功率.以在Yahoo!Answers数据集上训练的双向长短期记忆网络模型为例,用R-attack算法攻击模型的攻击成功率比基线算法高了2.4%.
A Black Box Algorithm of Random Beam Search Text Attack
To solve the problem that existing adversarial text generation algorithms are prone to fall into local optimal solution, an algorithm R-attack is proposed that uses beam search and random elements to improve the attack success rate.The R-attack algorithm first utilizes beam search to thoroughly explore the synonym space, thereby increasing the diversity of adversarial samples and enhancing the efficiency of the attack.Meanwhile, during the iterative search process, random elements are introduced to avoid premature convergence to local optima, effectively improving the success rate of the attack.Adversarial attack experiments were conducted on two models across three datasets, and the results demonstrate that the R-attack algorithm significantly improves the success rate of adversarial samples.Taking the example of attacking an LSTM model trained on"Yahoo! Answers,"the R-attack algorithm achieves a 2.4% increase in attack success rate compared to the baseline.

adversarial attacknatural language processingblack box attack

王小萌、张华、丁金扣、王稼慧

展开 >

北京邮电大学 网络与交换技术国家重点实验室,北京100876

对抗攻击算法 自然语言处理 黑盒攻击

国家自然科学基金

62072051

2024

北京邮电大学学报
北京邮电大学

北京邮电大学学报

CSTPCD北大核心
影响因子:0.592
ISSN:1007-5321
年,卷(期):2024.47(2)
  • 6