基于自优化深度网络的模型攻击方法
2021-11-09吴吉王月娟景栋盛
吴吉 王月娟 景栋盛



摘 要:机器学习方法常使用私有数据来训练模型以期获得更好的效果。然而,非授权用户可以通过模型输出来判断数据是否参与训练,破坏了数据隐私安全。对此,提出了基于深度优化网络的模型攻击方法,从攻击者的角度出发,分析攻击方法原理,有针对性地防御对模型的攻击,增强模型的隐秘性。所提方法自动对模型进行攻击,获得自优化的参数,提高攻击的准确度,充分挖掘模型中的安全缺陷,揭示模型的可改进之处,改善模型的安全性。在CIFAR-100数据集上进行了实验,得到AUC值为0.83,优于base方法。实验结果验证该方法能有效地提升攻击效果。
关键词:机器学习;优化;隐私保护;模型攻击
中图分类号:TP309 文献标识码:A
A Model Attack Method based on Self-optimizing Deep Network
WU Ji, WANG Yuejuan, JING Dongsheng
(Suzhou Power Supply Branch, State Grid Jiangsu Electric Power Co., Ltd., Suzhou 215004, China)
13862159678@163.com; 215691852@qq.com; jds19810119@163.com
Abstract: Machine learning often uses private data to train model so as to get better performance. However, unauthorized users can input data into the model and determine whether certain data are used for training by the output of the model, which threatens data privacy and security. In order to solve this problem, this paper proposes an attack method based on deep optimizing network, which analyzes the attack method principle from the attacker's point of view, and then defends against the attack on the model in a targeted manner so as to enhance the secrecy of the model. The proposed method attacks the model automatically, obtains self-optimizing parameters, improves the attack accuracy, fully exploits security defects in the model, reveals the improvement of the model, and improves the model security. Experiments have been carried out on CIFAR-100 data set, and the AUC (Area Under the Curve) value is 0.83, which is better than the base method. Experimental results show that the proposed method can effectively improve the attack effect.
Keywords: machine learning; optimization; privacy protection; model attack
1 引言(Introduction)
随着深度学习研究的不断深入,深度学习模型的安全问题引起了研究者的广泛关注,隐私泄露问题越来越受到重视[1-4]。一方面,模型固有的特性使攻擊者有机会获取其中的隐私信息;另一方面,模型中的隐藏层会形成较大的有效容量,将一些训练数据细节化为参数[5],记录在模型中。
通过对测试数据的输出分析,可以对模型有一个明确的衡量,同时也急需一个有效的攻击方法来模拟对目标的攻击,发现模型中存在的问题。虽然已有一些方法,但是这些方法在模型的攻击精度上还有待提高。因此,需要设计研发一种有效的方法来提高攻击的效果,从而更好地改进模型的安全性。
针对这一问题,本文提出了基于自优化的深度网络模型攻击方法,通过已知模型的层数,对其进行计算,得出一组攻击参数,使用这些指定参数对模型相应的层进行攻击,获得较好的攻击效果。……
