Core Concepts
フェデレーテッドラーニングにおけるプライバシー保護とモデル精度のトレードオフを最適化するための証明可能な相互利益を研究しました。
Stats
ˆµi has the following property E((ˆµi − µ)2) = 1 / (γi + ρ).
εi ≤ καi, where κ = 16√(2e ln(1.25n^2) ln(4n^2)B / n).
The outcome ∆wm := E(||wm − w∗||^2) satisfies ∆wm ≤ 1 / (1+χ/2−χ(m−T))LµΓ.
Quotes
"Such a delicate balance between model accuracy and data privacy brings into question the viability of FL."
"Mutually beneficial protocol ensures that all clients will benefit from collaborative learning compared to individual learning."
"In the limit N → ∞, it is profitable to collaborate."