Strategic dropout, explained
Availability is part of the problem. What if an adversary decides which clients go missing?
On the light side the diagram shows random dropout: a different set of clients misses each round. The dark side shows the attack from the IC3AI paper: the same high-value clients go missing every round, and thick lines show AAWA up-weighting their neighbours.
Random dropout. A different 6 of the 20 clients miss each round.
The reported result
Strategic suppression. An adversary removes the same 6 high-value clients every round.
Accuracy lost as suppression grows from 10 to 50 percent
- AAWA loses
- 5.94 points
- FedAvg loses
- 7.57 points
At 30 percent: FedAvg 37.34%, AAWA 35.45%.
Reported from single runs: when an adversary decides which clients go missing, aggregation that tracks who is absent loses accuracy more slowly than FedAvg. A pre-registered re-test found the gain indistinguishable from zero.