守门员:防御相关值和符号编码攻击
Rongfei Zhuang1, Ximing Fu1,2, Chuanyi Liu1,2
1School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen 518055, China.
Entropy (Basel, Switzerland)
|March 28, 2025
概括
Goalie通过早期检测恶意模型来保护数据信托免受数据盗窃攻击. 这种方法通过阻止有害的训练来保护良性模型的性能,与现有的防御不同.
科学领域:
- 机器学习安全 机器学习安全
- 数据信任安全数据信任安全
- 保护隐私的机器学习
背景情况:
- 数据信托面临相关价值和符号编码攻击来窃取共享数据.
- 当前防御通过扰乱参数或数据来降低良性模型性能.
- 存在对保护数据而不损害模型实用性的方法的需求.
研究的目的:
- 提出Goalie,一种新的方法来保护数据信托免受数据盗窃攻击.
- 检测和阻止恶意模型训练,同时保持良性模型性能.
- 确保数据信托中共享数据的完整性和安全性.
主要方法:
- Goalie通过分析编码攻击引起的参数分布变化来检测恶意模型.
- 在早期培训时代,从模型参数中提取特征.
- 理论分析证实,攻击中使用的规范化术语会改变参数分布.
主要成果:
- 在检测具有编码攻击的模型时,Goalie 实现了高精度 (超过 0.9).
- 该方法即使在极端攻击场景中也表现出有效性.
- 检测效率很高,使用早期时代特征只需要1.1毫秒.
结论:
- 目标有效地捍卫数据信任对复杂的数据盗窃攻击.
- 拟议的检测机制保留了良性模型的性能.
- 戈利为数据信任安全提供了一个计算效率高,准确的解决方案.
相关概念视频
Insulation Coordination
Insulation coordination is the process of matching electric equipment's insulation strength with protective device characteristics to protect the equipment against expected overvoltages. This selection is based on engineering judgment and cost. Equipment can generally withstand short-duration high transient overvoltages, but repeated tests with identical waveforms can yield inconsistent results. As a result, standard impulse voltage waveforms are used for testing, defined by specific times for...
Radial System Protection
Radial systems employ time-delay overcurrent relays to reduce load interruptions. When a fault occurs, the nearest breaker opens first, while upstream breakers remain closed due to longer delay settings. This approach ensures minimal disruption to the rest of the system.
In a radial system with a fault downstream of the third breaker, ideally, only the third breaker will open, isolating the fault and interrupting the load connected beyond it. The second breaker has a longer delay setting,...
In a radial system with a fault downstream of the third breaker, ideally, only the third breaker will open, isolating the fault and interrupting the load connected beyond it. The second breaker has a longer delay setting,...
Zones of Protection
In power systems, the entire setup is divided into protective zones to isolate faults and protect the rest of the network. These zones include generators, transformers, buses, transmission lines, distribution lines, and motors. Each zone can be visualized as a separate room in a house, with each room protected by its own circuit breaker.
Protective zones are defined by closed dashed lines, containing one or more components. A key characteristic of these zones is the strategic placement of...
Protective zones are defined by closed dashed lines, containing one or more components. A key characteristic of these zones is the strategic placement of...
Line Protection with Impedance Relays
Coordinating time-delay overcurrent relays in complex radial systems and directional overcurrent relays in multi-source transmission loops can be challenging. Impedance relays address these issues by responding to the voltage-to-current ratio, specifically measuring the apparent impedance of a line. These relays become more sensitive during faults as current increases and voltage decreases, thereby reducing the apparent impedance.
Under normal conditions, low load currents keep the measured...
Under normal conditions, low load currents keep the measured...
Pilot and Numeric Relaying
Pilot relaying is a type of differential protection used in power systems. It compares electrical quantities at the terminals of equipment via a communication channel instead of direct relay interconnection. This method is essential for transmission lines where the terminals are far apart, typically up to 80 km for lines with 69 to 115 kV ratings. Four types of communication channels are used for pilot relaying:
Censoring Survival Data
Survival analysis is a statistical method used to analyze time-to-event data, often employed in fields such as medicine, engineering, and social sciences. One of the key challenges in survival analysis is dealing with incomplete data, a phenomenon known as "censoring." Censoring occurs when the event of interest (such as death, relapse, or system failure) has not occurred for some individuals by the end of the study period or is otherwise unobservable, and it might have many different reasons...


