Poisoning Knowledge Graph Embeddings via Relation Inference Patterns

TL;DR

This paper introduces a novel data poisoning attack on knowledge graph embedding models by exploiting their relation inference patterns, significantly degrading link prediction performance.

Contribution

It proposes a new poisoning method leveraging symmetry, inversion, and composition patterns to craft adversarial facts that mislead KGE models.

Findings

Poisoning attacks outperform baselines on multiple models and datasets.

Symmetry-based attacks are highly effective and generalize across models.

KGE models are sensitive to relation inference patterns, especially symmetry.

Abstract

We study the problem of generating data poisoning attacks against Knowledge Graph Embedding (KGE) models for the task of link prediction in knowledge graphs. To poison KGE models, we propose to exploit their inductive abilities which are captured through the relationship patterns like symmetry, inversion and composition in the knowledge graph. Specifically, to degrade the model's prediction confidence on target facts, we propose to improve the model's prediction confidence on a set of decoy facts. Thus, we craft adversarial additions that can improve the model's prediction confidence on decoy facts through different inference patterns. Our experiments demonstrate that the proposed poisoning attacks outperform state-of-art baselines on four KGE models for two publicly available datasets. We also find that the symmetry pattern based attacks generalize across all model-dataset combinations which indicates the sensitivity of KGE models to this pattern.