Lecture 19

Federated Learning

image.png

Deep Leakage from Gradients

Gradients Matching

image.png

Defense Strategy

  1. Gaussian and Laplacian Noise
    1. ⇒ 효과 X!
      1. Noise 증가 → Leakage 감소 하지만 Acc도 감소 ㅠㅠ
  2. Gradient Compression (Pruning)
    1. ⇒ 효과 O!

      1. 30% 이상 prune했을 때 no leak!

      image.png

Memory bottleneck of on-device training

CNN

TinyTL

Tiny Transfer Learning

image.png