分散式訓練梯度傳輸污染

Risk Description

In distributed training environments, attackers compromise a training node to inject malicious gradient updates during transmission. Since distributed systems rely on gradient aggregation across multiple nodes, poisoned gradients affect global model parameters during aggregation. This attack is difficult to detect as anomalous gradients from a single node may appear as normal statistical variance.