Computer Vision Foundations · 免费预览
隐私、伦理与部署
去标识化数据如何被重新识别、差分隐私如何保护参与者，以及那些在每一次技术修复之后依然存在的伦理问题。
1. Anonymous data is not private data
Anonymous data is not private data
2. Differential privacy
Differential privacy
3. Randomized response
Randomized response
4. Concerns beyond bias
Concerns beyond bias
5. Questions to carry into practice
Questions to carry into practice
6. Recovering the rate from noisy answers
A survey uses the randomized response protocol: answer truthfully on the first coin's heads, and answer the second coin otherwise. Across 1000 participants, 400 stored answers are yes. Estimate the true fraction of yes answers in the population, and write it as a decimal.
7. What the re-identification case teaches
In the 1997 Massachusetts case, "anonymized" hospital records were linked to the governor's medical history. Which lesson about privacy follows for vision datasets?
8. Automation bias at deployment
A deployed vision system mislabels an item, and a human reviewer is available to catch such mistakes. Why does automation bias make this safety net weaker than it sounds?