STDatav2:访问有效的黑盒窃取对敌对攻击的攻击
概括
这项研究通过使用合成和代理数据来增强代孕模型训练,以防止模式崩. 一个新的自我条件框架确保了数据多样性和特定类别的约束,以改善黑盒模型窃取防御.
科学领域:
- 机器学习 机器学习
- 人工智能的人工智能
- 网络安全 网络安全
背景情况:
- 由于极端的设置,在没有训练数据的情况下窃取黑子模型具有挑战性.
- 像替代训练数据 (STDatav1) 这样的现有方法面临的局限性包括潜在的模式崩和维护数据多样性.
研究的目的:
- 为了提高代孕模型培训黑盒模型盗窃的有效性.
- 为了减轻模式崩,增强数据多样性和替代数据生成中的类特定约束.
主要方法:
- 提出了一个联合数据优化方案,使用合成和代理数据来训练代用模型.
- 引入了自有条件的数据合成框架,用于类特定的约束和多样性的伪类映射.
- 整合了STDatav1的类特定约束,并设计了双交叉损失函数.
主要成果:
- 与之前的STDatav1.1相比,表现出相当大的性能提升.
- 使用八种模型对四个数据集的评估证实了该方法的竞争能力和潜力.
- 提出的方法有效地解决了模式崩,并保持了数据多样性和类特定约束.
结论:
- 增强的代孕培训方法为黑盒模型盗窃提供了显著的改进.
- 联合数据优化和自我条件合成框架代表了该领域的一个有前途的进步.
- 这项工作提供了一种强大的方法来生成高质量的替代数据,推进模型隐私和安全方面的研究.
相关概念视频
Stereotype Content Model
13.9K
The Stereotype Content Model (SCM) was first proposed by Susan Fiske and her colleagues (Fiske, Cuddy, Glick & Xu, 2002; see also Fiske, 2012 and Fiske, 2017). The SCM specifies that when someone encounters a new group, they will stereotype them based on two metrics: warmth—or that group’s perceived intent, and how likely they are to provide help or inflict harm—and competence—or their ability to carry out that objective. Depending on the warmth-competence...
13.9K
Blinding
2.4K
Blinding is a commonly used method of not telling participants which treatment a subject is receiving. Blinding is a critical part of a randomized control trial or RCT. It reduces the bias that affects the results. In an RCT, blinding is used in the form of a placebo. A placebo effect occurs when untreated subjects falsely believe they have received the treatment and report improved symptoms. A placebo or a dummy treatment is administered to subjects to negate the bias caused by such an effect.
2.4K
Law of Effect
1.3K
B.F. Skinner, a prominent figure in behavioral psychology, introduced operant conditioning by emphasizing the role of consequences in shaping behavior. This theory builds upon the law of effect proposed by Edward Thorndike, which posits that behaviors followed by satisfying outcomes are likely to be repeated. In contrast, those followed by unsatisfying outcomes are less likely to recur.
Edward Thorndike's foundational work involved studying learning in animals, particularly using puzzle...
Edward Thorndike's foundational work involved studying learning in animals, particularly using puzzle...
1.3K


