Unmasking Diluted Triggers: An Explainable Framework for Robust Backdoor Detection in Long-Context NLP

Pre-trained Language Models (PLMs) have become foundational in Natural Language Processing, yet they remain vulnerable to backdoor attacks where hidden triggers manipulate model behavior. Existing defenses, particularly perturbationbased methods (e.g., STRIP) and perplexity-based approaches (e.g., ONION), exhibit critical failures when facing "Signal Dilution" in long-context documents, often degrading to randomguess performance. Furthermore, these methods typically operate as black boxes, lacking interpretability. To address these challenges, we propose D XAI, an explainable backdoor detection framework that unmasks triggers by analyzing the model's internal attribution landscape. Leveraging the proposed Funnel Effect mechanism-where triggers monopolize gradient flow-D XAI fuses Attribution Energy (signal intensity) and Distribution Inequality (signal sparsity) to isolate the distinctive signature of backdoor triggers. Extensive experiments across four benchmarks (SST-2, AG News, IMDb, and Thai Wongnai) demonstrate D XAI's superiority, achieving robust detection (AUROC ≈ 1.00) on long-context samples where baselines fail. Moreover, D XAI offers a computationally efficient inference capability, operating over 10× faster than perturbation-based counterparts, establishing it as a transparent and practical defense for deployed PLMs.