我對合成數據本質的三個思考：可信任元素，任務導向生成模型，基準真相。

思考#1：合成數據，始於隱私，終於更多的可信任AI元素。

1. 合成數據，是有顯著前景的科技。(Synthetic data is a technology with significant promise.)
2. 合成數據，可以加速發展。(Synthetic data can accelerate development.)
3. 合成數據，並非天生就有隱私性。(Synthetic data is not automatically private.)
4. 合成數據，並非真實數據的替代。(Synthetic data is not a replacement for real data.)
5. 離群值很難用保護隱私的方式抓到。(Outliers are hard to capture privately.)
6. 只用實驗評估單個數據集的隱私，是很有問題的。(Empirically evaluating the privacy of a single dataset can be problematic.)
7. 當生成合成數據，黑盒子模型特別不透明。(Black box models can be particularly opaque when it comes to generating synthetic data.)
8. 合成數據不只能幫助隱私。(Synthetic data goes beyond privacy.)

思考#2：分析合成數據的三個維度，資料科學任務，目的導向模型，數據生成方法

Definition 1 Synthetic data is data that has been generated using a purpose-built mathematical model or algorithm, with the aim of solving a (set of ) data science task(s).

1. 資料科學任務 (Data Science Tasks)
2. 目的建築數學模型或演算法 (Purpose-built mathematical model or algorithm)
3. 生成的數據 (Generated Data)

思考#3：合成數據，是基準真相，也迴避隱私顧慮

• 利用電腦生成合成數據(Computer-generated synthetic data)來解任務，最找可以追到Ulam與von Neumann在1940s年代的蒙地卡羅模擬方法(Monte Carlo simulation methods)。
• 模擬方法在研究常被用，因為他能夠給一種基準真相 (Ground truth)，來幫助人去發展與衡量機器學習途徑。
• 現在研究合成數據的驅動，是來自「數據保護法規(Data Protection Regulations)」，希望能用合成數據，來迴避隱私顧慮。

• Train on Synthetic, Test on Real (TSTR): https://arxiv.org/pdf/1706.02633.pdf
• Hide-and-Seek Privacy Challenge: Synthetic Data Generation vs. Patient Re-identification. http://proceedings.mlr.press/v133/jordon21a/jordon21a.pdf
• Pate-gan: Generating synthetic data with differential privacy guarantees. https://openreview.net/pdf?id=S1zk9iRqF7

Utility: The utility of synthetic data often is determined by its usefulness for a given task or set of tasks. This often involves contrasting the performance of models trained on real vs synthetic data, and might involve inspecting concrete metrics such as accuracy, precision, root mean-squared error, etc.; and/or model fairness properties such as demographic parity, fairness through unawareness, or conditional fairness [23]. Doing so often requires the Train on Synthetic, Test on Real (TSTR) paradigm [16] in which models are trained on synthetic data and their performance then evaluated on real data