ShadowLogic: Backdoors in Any Whitebox LLM
"后门攻击" 相关文章
Stealthy Dual-Trigger Backdoors: Attacking Prompt Tuning in LM-Empowered Graph Foundation Models
TED++: Submanifold-Aware Backdoor Detection via Layerwise Tubular-Neighbourhood Screening
TED++: Submanifold-Aware Backdoor Detection via Layerwise Tubular-Neighbourhood Screening
Injection, Attack and Erasure: Revocable Backdoor Attacks via Machine Unlearning
只要250份投毒樣本，就能在LLM預訓練植入後門觸發亂碼
DemonAgent: Dynamically Encrypted Multi-Backdoor Implantation Attack on LLM-based Agent
DemonAgent: Dynamically Encrypted Multi-Backdoor Implantation Attack on LLM-based Agent
TabVLA: Targeted Backdoor Attacks on Vision-Language-Action Models
TabVLA: Targeted Backdoor Attacks on Vision-Language-Action Models
TabVLA: Targeted Backdoor Attacks on Vision-Language-Action Models
Rounding-Guided Backdoor Injection in Deep Learning Model Quantization
250份恶意样本“毒倒”百亿参数大模型？内容安全已成AI生命线
250份恶意样本“毒倒”百亿参数大模型？内容安全已成AI生命线
250份恶意样本“毒倒”百亿参数大模型？内容安全已成AI生命线
250份文档就能给大模型植入后门：不分参数规模
250份文档就能给大模型植入后门：不分参数规模
Anthropic 发现：仅 250 份中毒文件即可攻破大型 AI 模型
管你模型多大，250份有毒文档统统放倒，Anthropic：LLM比想象中脆弱
管你模型多大，250份有毒文档统统放倒，Anthropic：LLM比想象中脆弱
