On the Hidden Costs of Counterfactual Knowledge Training in LLM Unlearning
该论文研究了大型语言模型(LLM)概念消除中的反事实微调(CFT)范式。CFT通过训练模型生成替代的虚构知识来替代不需要的内容,是一种有前景的方法。然而,作者发现CFT在某些方面仍不如其他范式,并揭示了两个被忽视的问题:(1)知识冲突:反事
今日共收录 1142 条安全情报,包含 375 个 CVE,381 篇安全通告,以及 103 篇研究论文。
← 返回归档列表该论文研究了大型语言模型(LLM)概念消除中的反事实微调(CFT)范式。CFT通过训练模型生成替代的虚构知识来替代不需要的内容,是一种有前景的方法。然而,作者发现CFT在某些方面仍不如其他范式,并揭示了两个被忽视的问题:(1)知识冲突:反事
该论文针对开放权重大语言模型(LLM)的微调防护机制进行了系统性研究。现有防护假设有害行为只能通过微调习得,但论文指出预训练模型已蕴含大量有害知识,因此攻击者可能无需微调,仅通过越狱手段即可绕过防护。作者评估了两种低成本攻击方法:ablit
CrowdStrike Named a Leader in Frost & Sullivan 2026 Radar for Cloud-Native Application Protection Platforms
CrowdStrike Expands Real-Time Cloud Detection and Response to Google Cloud
CrowdStrike Falcon Cloud Security Delivered 264% ROI Through Unified Cloud Protection
Security & Identity
SAP on Google Cloud
Inside Google Cloud
Google Cloud Next & Events
Google Cloud Consulting
Transform with Google Cloud
Threat IntelligenceNorth Korea-Nexus Threat Actor Compromises Widely Used Axios NPM Package in Supply Chain AttackBy Goo
Google Cloud Products
Cloud & Application Security
From Scanner to Stealer: Inside the trivy-action Supply Chain Compromise
Next-Gen Identity Security
CrowdStrike FalconID Brings Phishing-Resistant MFA to Falcon Next-Gen Identity Security
CrowdStrike Named a Leader in Identity Threat Detection and Response
Nimbus Manticore has continued its operations during and after the US military campaign against Iran. The post Iranian A
FuzzPilot 是一个针对 AFL++ 模糊测试工具的控制器,旨在解决结构化文本(如 JSON、XML)模糊测试中的覆盖率平台期问题。传统上,AFL++ 的突变热路径中会执行昂贵的推理操作(如语法分析、语义验证),导致性能瓶颈。Fuzz