signed bit flip 最近有什么新动态？

traeai 已收录 1 篇与 signed bit flip 相关的内容。最新一篇是「AI Models May Be More Fragile Than We Think」，由 Last Week in AI 发布。

概念

signed bit flip

别名：符号位翻转

针对浮点数符号位的攻击操作，将正数变为负数或反之，改变数值方向而不改变量级。

已跟踪 1 条高相关材料

TraeAI 观察

如果只读 3 篇

AI Models May Be More Fragile Than We Think

Last Week in AI · 7.5 分

仅需翻转 ResNet-50 中两个参数的符号位（signed bit flip），即可在无需数据访问或训练的情况下使 ImageNet 准确率暴跌 99.8%，揭示神经网络存在极端脆弱的"深层病灶"攻击面。

AI Models May Be More Fragile Than We Think

Last Week in AI5月19日182 字 (约 1 分钟)

Flipping just two signed bits in ResNet-50 can cause a 99.8% drop in ImageNet accuracy without any data access or training, revealing an extreme vulnerability called 'deep neural lesion' in neural networks.

入选理由：攻击仅需翻转 2 个参数的符号位，无需接触训练数据或进行优化

FeaturedVideo#Neural Network Security#Adversarial Attack#Model Robustness#Bit-level Attack#ResNet英文

跨材料问答 · signed bit flip

回答基于：signed bit flip 相关 1 条材料