对抗性例子和稳健性测验

当您参加测验时，就会出现答案和解释。

1. Under a specified threat model, what is an adversarial example?
   • An input crafted to cause a model error using changes allowed by that threat model
   • Any ordinary prediction error on unmodified input
   • An input that must contain tiny, imperceptible pixel changes
   • A model update that improves accuracy on every distribution
2. “白盒”攻击与“黑盒”攻击有何区别？
   • 白盒攻击仅适用于白色图像
   • 在白盒中，攻击者可以访问模型的内部/梯度；在黑盒中，仅输出
   • 黑盒攻击总是更快
   • 白盒攻击不需要计算
3. FGSM（快速梯度符号法）通过...制作了一个对抗性示例
   • 随机打乱像素
   • 从头开始重新训练模型
   • 相对于输入，沿着损失梯度的符号方向迈出一步
   • 裁剪图像
4. What is adversarial training commonly used for?
   • Guaranteeing robustness to every possible attack
   • Proving a classifier is fair
   • Eliminating all clean-data errors
   • Improving robustness against the adversarial threat model represented during training
5. 为什么对抗性例子的“可转移性”值得关注？
   • 针对一种模型的攻击通常会愚弄其他模型，从而实现黑盒攻击
   • 它使模型训练速度变慢
   • 它增加了存储成本
   • 它仅影响音频模型
6. How are prompt-injection attacks related to image adversarial examples?
   • They are identical bounded pixel perturbations
   • Both exploit model vulnerabilities, but prompt injection has a distinct instruction and system threat model
   • Prompt injection requires changing model weights
   • Image adversarial examples only affect text prompts
7. Why must an adversarial-robustness report specify the perturbation norm and budget?
   • They determine the model’s number of parameters
   • They guarantee transfer to every attack
   • They define which input changes the reported robustness claim covers
   • They replace the need to state an attacker’s access
8. Which pair of measures helps distinguish clean performance from performance under a specified attack?
   • Training-set size and GPU memory
   • Tokenization rate and context length
   • Model file size and parameter dtype
   • Clean accuracy and robust accuracy under the stated attack setting