BackX

BackX是一个基于后门攻击的解释性AI基准数据集，用于评估深度模型预测的归属方法。该数据集由西澳大学创建，旨在通过理论分析确保归属评估的可靠性。BackX通过控制特定的触发模式来系统地建立模型的归属真相，从而在不同底层方法中实现归属方法的一致和公平基准。该数据集的应用领域包括解释模型预测和建立模型信任，特别是在高风险任务中。BackX is a backdoor attack-based explainable AI benchmark dataset dedicated to evaluating attribution methods for deep model predictions. Developed by The University of Western Australia, this dataset aims to ensure the reliability of attribution evaluation via theoretical analysis. BackX systematically constructs the ground truth for model attribution by controlling specific trigger patterns, thereby enabling consistent and fair benchmarks for attribution methods across different underlying approaches. Its application scenarios cover explaining model predictions and establishing model trust, particularly in high-risk tasks.