14.6.4: Case Study Questions
- Page ID
- 58362
\( \newcommand{\vecs}[1]{\overset { \scriptstyle \rightharpoonup} {\mathbf{#1}} } \)
\( \newcommand{\vecd}[1]{\overset{-\!-\!\rightharpoonup}{\vphantom{a}\smash {#1}}} \)
\( \newcommand{\dsum}{\displaystyle\sum\limits} \)
\( \newcommand{\dint}{\displaystyle\int\limits} \)
\( \newcommand{\dlim}{\displaystyle\lim\limits} \)
\( \newcommand{\id}{\mathrm{id}}\) \( \newcommand{\Span}{\mathrm{span}}\)
( \newcommand{\kernel}{\mathrm{null}\,}\) \( \newcommand{\range}{\mathrm{range}\,}\)
\( \newcommand{\RealPart}{\mathrm{Re}}\) \( \newcommand{\ImaginaryPart}{\mathrm{Im}}\)
\( \newcommand{\Argument}{\mathrm{Arg}}\) \( \newcommand{\norm}[1]{\| #1 \|}\)
\( \newcommand{\inner}[2]{\langle #1, #2 \rangle}\)
\( \newcommand{\Span}{\mathrm{span}}\)
\( \newcommand{\id}{\mathrm{id}}\)
\( \newcommand{\Span}{\mathrm{span}}\)
\( \newcommand{\kernel}{\mathrm{null}\,}\)
\( \newcommand{\range}{\mathrm{range}\,}\)
\( \newcommand{\RealPart}{\mathrm{Re}}\)
\( \newcommand{\ImaginaryPart}{\mathrm{Im}}\)
\( \newcommand{\Argument}{\mathrm{Arg}}\)
\( \newcommand{\norm}[1]{\| #1 \|}\)
\( \newcommand{\inner}[2]{\langle #1, #2 \rangle}\)
\( \newcommand{\Span}{\mathrm{span}}\) \( \newcommand{\AA}{\unicode[.8,0]{x212B}}\)
\( \newcommand{\vectorA}[1]{\vec{#1}} % arrow\)
\( \newcommand{\vectorAt}[1]{\vec{\text{#1}}} % arrow\)
\( \newcommand{\vectorB}[1]{\overset { \scriptstyle \rightharpoonup} {\mathbf{#1}} } \)
\( \newcommand{\vectorC}[1]{\textbf{#1}} \)
\( \newcommand{\vectorD}[1]{\overrightarrow{#1}} \)
\( \newcommand{\vectorDt}[1]{\overrightarrow{\text{#1}}} \)
\( \newcommand{\vectE}[1]{\overset{-\!-\!\rightharpoonup}{\vphantom{a}\smash{\mathbf {#1}}}} \)
\( \newcommand{\vecs}[1]{\overset { \scriptstyle \rightharpoonup} {\mathbf{#1}} } \)
\(\newcommand{\longvect}{\overrightarrow}\)
\( \newcommand{\vecd}[1]{\overset{-\!-\!\rightharpoonup}{\vphantom{a}\smash {#1}}} \)
\(\newcommand{\ket}[1]{\left| #1 \right>}\)
\(\newcommand{\bra}[1]{\left< #1 \right|}\)
\(\newcommand{\braket}[2]{\left< #1 \vphantom{#2} \right| \left. #2 \vphantom{#1} \right>}\)
\(\newcommand{\braopket}[3]{\left< #1 \vphantom{#2}\vphantom{#3} \right| #2 \vphantom{#1}\vphantom{#3} \left| #3 \vphantom{#1}\vphantom{#2} \right>}\)
\(\newcommand{\qmvec}[1]{\mathbf{\vec{#1}}}\)
\(\newcommand{\op}[1]{\hat{\mathbf{#1}}}\)
\(\newcommand{\expect}[1]{\langle #1 \rangle}\)
\(\newcommand{\dfn}[1]{\emph{\textbf{#1}}}\)
Case Study Analysis
Google: Federated Learning for Privacy-Preserving Analytics
Google pioneered federated learning with its mobile keyboard (Gboard). The challenge: improve next-word prediction by learning from millions of users' typing without centralizing that data. The solution: each phone trains a small model update locally using only that user's data. Model updates (gradients, not raw data) are sent to Google's servers, aggregated across millions of devices, and used to improve the global model. The updated model is pushed back to all devices.
Key technical aspects: (1) Communication efficiency — updates are compressed to minimize bandwidth. (2) Differential privacy — noise is added to aggregated updates so individual contributions cannot be reverse-engineered. (3) Secure aggregation — a cryptographic protocol ensures the server only sees the aggregate, not individual updates.
Federated learning has since expanded to healthcare (hospitals training diagnostic models jointly without sharing patient records), finance (banks collaborating on fraud detection without sharing transaction data), and automotive (vehicles sharing driving data for autonomous driving improvements).
Discussion Questions:
-
Explain how federated learning solves the privacy-utility trade-off. Why is it better than simply anonymizing data and centralizing it? What are the limitations of anonymization that federated learning addresses?
-
Google adds differential privacy ON TOP of federated learning. Why isn't federated learning alone sufficient for privacy? What additional risk does differential privacy mitigate?
-
A consortium of 10 hospitals wants to use federated learning to jointly train a cancer detection model without sharing patient data. Describe the workflow: what happens on each hospital's servers, what is sent to the central coordinator, and how is the final model produced? What challenges might arise?
-
Federated learning adds complexity and can reduce model accuracy compared to centralized training. Under what circumstances is this accuracy trade-off acceptable? When should organizations insist on centralized data for better models?