Unverified50% confidenceFactExact time
拜占庭攻击者模型假设在 n 个参与方中至多有 f 个被恶意控制，且这些恶意方可以发送任意更新向量
1
Sources
50%
Confidence
Long-term
Relevance
8/19/2026
First Seen
Sources
抗投毒概念锚定：防御AI数据污染的新思路
Related Claims
Unverified攻击者可利用的推理链侧信道信号包括响应延迟、token概率分布（logprobs）和计费的token数量69% similarUnverifiedThreshold-based defenses face the threat of borderline stealthy poisoning attacks, where an attacker can keep each sample just at the edge of the tolerance threshold and erode concepts through cumulative effect.68% similarUnverified自动化对抗样本生成面临覆盖率幻觉挑战，自动生成的攻击样本可能数量庞大但高度同质化，遗漏新颖攻击路径66% similarUnverifiedResearchers can use n-gram overlap analysis and membership inference attacks to determine whether specific data appeared in a model's training set, but with the limitation of a high false positive rate.66% similarUnverified2024年研究揭示了多跳间接注入攻击，攻击者先污染模型记忆或工具调用状态，再通过后续交互完成数据外泄，使基于单次请求的检测系统失效65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/772861API
curl https://kongchang.com/api/v1/knowledge/claims/772861MCP
get_claim(id=772861)