We propose DIsoN, a decentralized method for out-of-distribution detection that enables a deployed model to directly compare incoming samples to the training data without data sharing.

For communication-efficient decentralized learning, we propose two new graphs: the k-peer exponential graph and the null-cascade graph.

We introduced and analyzed two novel gossip algorithms for rank and trimmed means estimation, proving convergence rates of $\mathcal{O}(1/t)$.