模型窃取攻击（model extraction）是什么？
What is a model extraction attack?
攻击者不断向一个只暴露预测接口的模型提问，根据返回的结果（如概率或标签）反推出模型的参数、结构或功能等价副本。它不需要接触训练数据，却可能侵犯知识产权并助长后续攻击。
An attacker queries a model that only exposes predictions and, from the returned labels or probabilities, reconstructs its parameters, architecture or a functionally equivalent copy—without touching training data, yet risking IP theft and further attacks.
难度 专业 · 适合 高中(15-18)
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#模型窃取 #攻击
什么是“数据偏见”，AI 为什么会带有偏见？
What is 'data bias' and why do AIs become biased?
如果训练数据本身就不均衡或有刻板倾向（例如某类人脸样本偏少），模型就会把这种偏差学进去并在预测中放大。所以 AI 的偏见往往反映的是训练数据与人类社会的既有偏见，而非凭空产生。
If training data is unbalanced or skewed (e.g. too few faces of some group), the model learns and amplifies that bias in its predictions. AI bias usually mirrors the bias already present in data and society, not something invented from nothing.
难度 进阶 · 适合 大众
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#偏见 #公平
对抗样本（adversarial example）是怎么“骗”过模型的？
How do adversarial examples fool a model?
攻击者在图片上叠加人眼几乎看不出的微小扰动，却让模型的内部激活大幅偏移，从而把熊猫认成吉他。这暴露了模型依赖的是人类不直观的脆弱特征，也提示自动驾驶等系统需要防备。
An attacker adds tiny, barely visible perturbations that shift the model's internals enough to label a panda as a guitar. It shows models rely on fragile, non-human features—a real concern for self-driving systems.
难度 进阶 · 适合 初中(12-15)
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#对抗样本 #安全
AI 会不会偷偷“学坏”？
Can an AI secretly 'learn bad things'?
AI 本身没有想法，它只是从你给的数据里学规律。如果数据里混进了错误、偏见或不良内容，它就会原样学进去并表现出来，所以“学坏”的根源通常在训练数据，而非它自己使坏。
AI has no intentions; it just learns patterns from the data you give. If that data contains errors, bias or harmful content, the AI learns and repeats them—so the root of 'bad learning' is usually the data, not the AI misbehaving on its own.
难度 入门 · 适合 小学低段(6-8)
CC BY 4.0 · 已核验 · Verified 2026-08-23 · 2 条独立证据
来源 · Source
Russell & Norvig,《Artificial Intelligence: A Modern Approach》第4版,Pearson,2020
S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020
独立核验依据 · Independent verification: 核验依据 2 ARCHIVE 标识 isbn_9788177583670：Artificial Intelligence: A
#AI #数据