# insider-threat
標記為「insider-threat」的 4 篇文章
Distributed Training Attack Surface
Security vulnerabilities in multi-GPU, multi-node LLM training: gradient sharing attacks, parameter server compromise, insider threats, and infrastructure-level training exploits.
distributed-traininggradient-sharingparameter-servermulti-gpuinsider-threatinfrastructure
Training Loop Vulnerabilities
Attacks on the training process itself including gradient manipulation, loss function tampering, learning rate schedule attacks, and training infrastructure compromise.
training-loopgradient-manipulationloss-functionlearning-rateinsider-threatsupply-chain
分散式訓練攻擊面
多 GPU、多節點 LLM 訓練中的安全漏洞：梯度共享攻擊、parameter server 入侵、內部威脅，以及基礎設施層級的訓練攻擊。
distributed-traininggradient-sharingparameter-servermulti-gpuinsider-threatinfrastructure
訓練迴圈漏洞
對訓練過程本身之攻擊，含梯度操弄、loss 函式篡改、學習率時程攻擊，與訓練基礎設施受損。
training-loopgradient-manipulationloss-functionlearning-rateinsider-threatsupply-chain