Recognizing Hereditary Properties in the Presence of Byzantine Nodes

TL;DR

This paper develops randomized algorithms for recognizing hereditary graph properties in distributed networks with Byzantine nodes, achieving near-optimal round complexity and establishing tight impossibility bounds.

Contribution

It introduces a new randomized approach for detecting hereditary properties in Byzantine settings with tight complexity bounds and proves the optimality of these results.

Findings

Algorithms work with high probability under Byzantine adversaries.

Round complexity is nearly optimal, matching lower bounds.

Recognizes properties like acyclicity, bipartiteness, with tight impossibility results.

Abstract

Augustine et al. [DISC 2022] initiated the study of distributed graph algorithms in the presence of Byzantine nodes in the congested clique model. In this model, there is a set $B$ of Byzantine nodes, where $|B|$ is less than a third of the total number of nodes. These nodes have complete knowledge of the network and the state of other nodes, and they conspire to alter the output of the system. The authors addressed the connectivity problem, showing that it is solvable under the promise that either the subgraph induced by the honest nodes is connected, or the graph has $2|B|+1$ connected components. In the current work, we continue the study of the Byzantine congested clique model by considering the recognition of other graph properties, specifically hereditary properties. A graph property is hereditary if it is closed under taking induced subgraphs. Examples of hereditary properties include acyclicity, bipartiteness, planarity, and bounded (chromatic, independence) number, etc. For each class of graphs ${\bf G}$ satisfying a hereditary property (a hereditary graph-class), we propose a randomized algorithm which, with high probability, (1) accepts if the input graph $G$ belongs to ${\bf G}$, and (2) rejects if $G$ contains at least $|B| + 1$ disjoint subgraphs not belonging to ${\bf G}$. The round complexity of our algorithm is $$O\left(\left(\dfrac{\log \left(\left|{\bf G}_n\right|\right)}{n} +|B|\right)\cdot\textrm{polylog}(n)\right),$$ where ${\bf G}_n$ is the set of $n$-node graphs in ${\bf G}$. Finally, we obtain an impossibility result that proves that our result is tight. Indeed, we consider the hereditary class of acyclic graphs, and we prove that there is no algorithm that can distinguish between a graph being acyclic and a graph having $|B|$ disjoint cycles.