paper implementation tracker

[tharvik]

over the years, a bunch of papers have been published at MLO and around which might help DISCO one way or the other. here goes a list of the ones we spoke about. it is still up for debate which we actually need the most.

• 🌪️ communication efficiency
  • PowerGossip
  • Choco-SGD
• 🔒 privacy-preserving
  • secure aggregation for privacy-preserving ML
    • somewhat done, only honest-but-curious server, need to be validated
  • secure byzantine-robust ML
• 🛠️ fault-tolerant and dynamic over time
  • buffered async aggregation
    • done, need to document it in code
  • async SGD
• 🥷 robust to malicious actors and data poisoning
  • byzantine robust optimisation
    • partially done via norm clipping
  • byzantine robust learning on heterogeneous data
• 🍎 🍌 interpretable in imperfectly interoperable data distributions
  • IFedAvg
• 🪞 personalizable
  • personalized FL

[martinjaggi]

only cosmetic change needed:

async buffered is already implemented.
byz as well (partially)

future roadmap: ... ifedavg
future roadmap: ... personalized
future roadmap: ... incentivize