Core Concepts
The authors leverage the dual approach to design a general robust decentralized optimization method, providing both global and local clipping rules in the special case of average consensus, with tight convergence guarantees. They also demonstrate how the clipping rules can serve as a basis for designing efficient attacks.
Abstract
The paper investigates Byzantine-resilient algorithms in a decentralized setting, where devices communicate directly with one another. The authors leverage the dual approach to design a general robust decentralized optimization method. They provide both global and local clipping rules in the special case of average consensus, with tight convergence guarantees.
For the global clipping rule, the authors show that it ensures the error decreases at each step, but cannot guarantee consensus among honest nodes. They construct an example where the honest nodes' models do not converge even with global clipping.
For the local clipping rule, the authors show that it guarantees linear convergence of the variance of honest nodes' parameters to 0. This ensures all honest nodes ultimately converge to the same model, though it may not be the optimal one due to the bias introduced by Byzantine corruption and the asymmetry of clipping.
The authors also show that local trimming (equivalent to Nearest Neighbor Averaging) is, up to a constant, as efficient as local clipping in sparse decentralized settings.
Finally, the authors propose a principled approach for designing attacks on communication networks, by exploiting the topology of the network.