Robust Truth Inference in Crowdsourcing under Adversarial Attacks via Graph Neural Networks

Abstract

Crowdsourcing is widely used to obtain labeled data for machine learning, but the openness of these platforms makes label quality vulnerable to noise and adversarial manipulation. Coordinated behaviors such as collusion and Sybil-style participation can create a consistent yet incorrect consensus, causing classical truth inference methods to amplify attacker-induced agreement and degrade downstream model performance. This paper proposes a graph neural network (GNN) framework for robust truth inference under adversarial attacks. We represent the worker–task ecosystem as a bipartite interaction graph in which edges correspond to submitted labels and node embeddings are learned via message passing, allowing the model to capture higherorder agreement and co-labeling structure. To avoid being dominated by correlated malicious behavior, the method incorporates task-content features for image and text tasks as an external semantic anchor that stabilizes learning when observed labels are heavily corrupted. The framework is trained with a joint objective that (i) predicts task labels and (ii) estimates worker trustworthiness, enabling simultaneous aggregation and risk-aware worker profiling. We evaluate the method in a controlled simulation setting on multiple datasets and attack strategies, including random labeling, single-class attacks, label flipping, and coordinated collusion. Experiments analyze label-aggregation accuracy, robustness under increasing adversarial participation, and malicious-worker detection under varying attack rates and task difficulty. Results indicate that combining relational signals with content anchoring improves robustness in adversarial label aggregation, especially under correlated attacks that target hard tasks, while producing conservative trust scores that support worker-risk scoring and downstream quality control. Overall, the proposed approach offers a unified graph-based solution for robust aggregation and adversary-aware monitoring in crowdsourced labeling workflows.