关于开放世界测试时间培训模型的稳定性研究
Shu Pi1, Xin Wang1, Jiatian Pi1
1National Center for Applied Mathematics In Chongqing, Chongqing Normal University, Chongqing, China.
Frontiers in artificial intelligence
|August 20, 2025
概括
新的测试时间毒害攻击利用了开放世界测试时间训练 (OWTTT) 模型中的漏洞. 这些攻击动态地产生对抗性干扰,损害了OWTTT的性能,并突显了对综合安全防御的需求.
科学领域:
- 人工智能
- 机器学习安全性
背景情况:
- 深度学习模型难以将其推广到新的领域,从而激发了测试时间培训/适应 (TTT/TTA).
- 开放世界TTT (OWTTT) 解决了分辨已退出分布的样本的挑战,但引入了新的安全风险,如测试时间中毒攻击.
- 测试时间中毒攻击与传统的不同,因为它们发生在模型的测试阶段,而不是训练.
研究的目的:
- 引入一种针对OWTTT模型的新型测试时间中毒攻击方法.
- 展示OWTT模型对这些新类型攻击的脆弱性.
- 强调在OWTT方法中需要强有力的安全措施.
主要方法:
- 开发了一种基于查询的单步方法,以动态生成和更新对抗性干扰.
- 在测试阶段利用动态变化的模型梯度.
- 在适应阶段对OWTTT模型产生了输入干扰.
主要成果:
- 拟议的测试时间中毒攻击严重损害了OWTTT模型的性能.
- 实验结果证实了在一个现实的OWTTT场景中攻击的有效性.
- 显示了当前OWTTT模型的重大漏洞.
结论:
- 没有安全评估的OWTT算法不适合在现实世界中部署.
- 有必要将对测试时间中毒攻击的防御整合到OWTTT方法中.
- 未来的OWTTT研究必须优先考虑安全与适应能力.
更多相关视频
07:01Working Memory Training for Older Participants: A Control Group Training Regimen and Initial Intellectual Functioning Assessment
Published on: September 20, 2020
4.8K
03:14Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
680
相关概念视频
Survival Tree
157
Survival trees are a non-parametric method used in survival analysis to model the relationship between a set of covariates and the time until an event of interest occurs, often referred to as the "time-to-event" or "survival time." This method is particularly useful when dealing with censored data, where the event has not occurred for some individuals by the end of the study period, or when the exact time of the event is unknown.
Building a Survival Tree
Constructing a...
Building a Survival Tree
Constructing a...
157
Wald-Wolfowitz Runs Test II
314
The Wald-Wolfowitz runs test, commonly referred to as the runs test, is a nonparametric test used to assess the randomness of ordered data. The test evaluates the number of runs, which are consecutive sequences of similar elements within the data. If the number of runs is significantly higher or lower than expected, the data is considered non-random, indicating a detectable pattern or structure.
For binary data, runs are identified using symbols such as + and −, or equivalently, 1s and...
For binary data, runs are identified using symbols such as + and −, or equivalently, 1s and...
314
Wald-Wolfowitz Runs Test I
738
The Wald-Wolfowitz test, also known as the runs test, is a nonparametric statistical test used to assess the randomness of a sequence of two different types of elements (e.g., positive/negative values, successes/failures). It examines whether the order of the elements in a sequence is random or if there is a pattern or trend present. This nonparametric test applies to any ordered data despite the population and sample data distribution, even if a higher sample size is available.
The test works...
The test works...
738
Three-Compartment Open Model
420
The three-compartment open model is a pharmacokinetic model used to describe the distribution and elimination of drugs following extravascular administration. It comprises a central compartment representing the plasma and two peripheral compartments. The highly perfused peripheral compartment represents organs and tissues with a rich blood supply, such as the liver, kidneys, and lungs. The scarcely perfused peripheral compartment represents tissues with lower blood supply, such as adipose...
420
Goodness-of-Fit Test
4.0K
The goodness-of-fit test is a type of hypothesis test which determines whether the data "fits" a particular distribution. For example, one may suspect that some anonymous data may fit a binomial distribution. A chi-square test (meaning the distribution for the hypothesis test is chi-square) can be used to determine if there is a fit. The null and alternative hypotheses may be written in sentences or stated as equations or inequalities. The test statistic for a goodness-of-fit test is given as...
4.0K
Observational Learning
305
Albert Bandura's observational learning, also known as imitation or modeling, occurs when a person observes and imitates another's behavior. It is a quicker process than operant conditioning. A well-known example is the Bobo doll study, where children who saw an adult acting aggressively towards the doll were more likely to act aggressively when left alone, compared to those who observed a nonaggressive adult. Many psychologists view observational learning as a form of latent learning...
305
