A Backdoor-based Explainable AI Benchmark for High Fidelity Evaluation of Attributions

Attribution methods need better testing tools: The paper addresses a critical gap in responsible AI—we currently lack reliable ways to verify whether explanation methods actually identify the features that truly influence a model's decisions. This matters because if we can't trust explanations, we can't trust AI systems to be transparent.

BackX provides a new benchmark with "ground truth": The researchers developed BackX, a testing framework that uses intentionally planted "backdoors" (hidden patterns) in AI models to create known explanations. This allows researchers to measure whether attribution methods correctly identify these planted features, providing the first reliable way to evaluate explanation accuracy.

Existing benchmarks have hidden problems: The paper identifies specific fidelity criteria that good benchmarks should meet and shows that current evaluation methods fall short. By establishing standardized testing conditions and controlling for confounding factors, BackX enables fairer comparisons between different explanation methods than previously possible.

Better explanations can improve AI security: Beyond improving transparency, the research reveals that attribution methods can help defend against "neural Trojans"—malicious hidden triggers inserted into AI models. Understanding which features matter most helps identify when models have been compromised, adding a security dimension to explainability work.