XAI-Grounded Explanation Generation for Speech Deepfake Detection with Training-Free Multimodal Large Language Models

2026-06-14

AI Summary

This work addresses the lack of trustworthy, interpretable, and decision-aligned explanations in current speech deepfake detection systems. The authors propose a training-free explanation framework that, for the first time, leverages low-level attribution signals generated by gradient-based explainable AI (XAI) methods as heuristic evidence to prompt a multimodal large language model, thereby producing natural language explanations that are evidence-based, specific, and easily understandable. The approach is evaluated on a newly constructed grounded explanation dataset derived from the PartialSpoof benchmark. Human evaluations and faithfulness tests demonstrate that incorporating XAI-derived evidence improves the internal accuracy of explanations by over 45%, enabling the generation of high-quality, task-specific, and verifiable explanations.

Abstract

Speech deepfake detection (SDD) systems require trustworthy explanations for reliable decision-making. Existing explanation ways mainly fall into two categories. Traditional explainable AI (XAI), such as gradient-based attribution, produces low-level attribution signals tightly coupled with model decisions, and harder to be understood by human than natural language explanations. Meanwhile, large language model (LLM)-based explanation generation often produces generic and ungrounded descriptions due to the lack of heuristic evidence and task-specific supervision, stemming from limited grounded explanation datasets for SDD. We therefore propose a training-free explanation framework that integrates XAI evidence with multimodal LLMs to generate grounded and specific explanations. Using the PartialSpoof dataset, we construct a grounded explanation dataset and show that methods with XAI increase inside accuracy by over 45\%, verified through human evaluation and faithfulness checks.

Problem

• Speech Deepfake Detection
• Explainable AI
• Large Language Models
• Grounded Explanations
• Multimodal Explanation

Innovation

• Explainable AI (XAI)
• Multimodal Large Language Models
• Speech Deepfake Detection
• Training-Free Explanation
• Grounded Explanation