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

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

Problem Statement
Speech Deepfake Detection (SDD) systems currently face a critical challenge in providing trustworthy, interpretable explanations for their decisions. While existing approaches utilize traditional Explainable AI (XAI) techniques (e.g., gradient-based attribution), these methods often produce low-level signals (such as saliency maps) that are tightly coupled with model internals and difficult for humans to interpret. Conversely, explanations generated by Large Language Models (LLMs) tend to be generic, ungrounded, and prone to hallucination due to a lack of heuristic evidence and task-specific supervision. Furthermore, there is a notable absence of dedicated datasets tailored for SDD explanations, and current LLM-based pipelines lack principled mechanisms to verify the faithfulness of their outputs.

Methodology: XAI-Grounded Explanation Generation (XGEG)
The authors propose a training-free framework, XGEG, which integrates traditional XAI evidence with Multimodal Large Language Models (MLLMs) to generate grounded, specific, and trustworthy explanations. The pipeline operates as follows:

1. Foundation Models & XAI Evidence Generation: Three pre-trained SDD models (based on wav2vec 2.0, HuBERT, and WavLM) are utilized without fine-tuning. These models generate attribution maps using Integrated Gradients (IG), Saliency, and LIME. Additionally, SHAP-derived importance scores are calculated for acoustic features extracted via the openSMILE eGeMAPSv02 set using a lightweight, pre-trained MLP classifier.
2. Data Filtering: The framework operates on the PartialSpoof dataset. To ensure reliability, only samples correctly classified by all four models (three foundation models + the MLP) are retained. The focus is strictly on spoofed samples, as explanations for bona fide speech are inherently less informative.
3. Multimodal Integration:
   • Visual Interpretation: Spectrogram-based heatmaps from IG, LIME, and Saliency are fed into a vision-LLM (Qwen2.5-VL-7B) to summarize abnormal regions in terms of temporal and frequency ranges.
   • Textual Synthesis: These time-frequency summaries, combined with top-ranked acoustic features from SHAP, are provided as input to a multimodal LLM (Qwen3-Omni-30B).
   • Prompting Strategy: The LLM is explicitly instructed to treat XAI evidence as supporting signals rather than definitive conclusions, prioritizing the acoustic content of the input. The model is prompted to critically analyze the evidence rather than passively reproducing it.
4. Structured Output: To mitigate hallucination, the output is constrained to a predefined structured format containing: (I) Audio Abnormality (Time/Freq Range), (II) Free-text Explanation, and (III) XAI Aggregation (indicating cross-model agreement).

Key Contributions

1. XAI-Grounded Framework: The paper introduces a principled framework that aggregates cross-model XAI signals and employs fidelity-driven validation to guide MLLMs. This approach produces semantically rich, specific explanations with reduced hallucination compared to purely text-based or single-model attribution methods.
2. Dataset Creation: The authors constructed and publicly released a large-scale explainable SDD dataset based on PartialSpoof, comprising approximately 65,000 explanation instances. This dataset addresses the gap in dedicated resources for explainable SDD research.
3. Training-Free Scalability: The methodology relies on pre-trained models and a training-free pipeline, ensuring scalability across diverse dataset generation settings without the need to retrain dedicated detectors for each task.

Results and Evaluation
The framework was evaluated through qualitative analysis, human evaluation, and quantitative metrics:

• Human Evaluation: Annotators rated explanations on correctness, evidence support, specificity, and overall preference. Results indicated that XAI-guided versions significantly outperformed the "Pure Audio" baseline. Notably, the All XAI (Three Model) configuration achieved the highest overall preference (1.50 average selection) and high specificity (4.30), demonstrating that multi-source XAI guidance reduces hallucinations and improves detail.
• Quantitative Localization:
  • Inside Accuracy (IA): The LIME single-method configuration achieved the highest IA (0.811), indicating superior precision in localizing abnormal time intervals without over-expansion.
  • Area-Normalised Local Logit Sensitivity: This metric, designed to test fidelity without destructive masking, showed that multimodal methods identify more influential regions than the baseline. The All XAI (Single Model) configuration achieved the highest sensitivity density (3.1466 × 10⁻⁶), representing a 22.10× improvement over the Pure Audio baseline.
• Bona Fide Analysis: The system was also tested on genuine audio, successfully generating explanations that justify the absence of anomalies, though the authors note this remains more challenging than detecting fakes.

Significance and Claims
The paper claims that the development of trustworthy and interpretable explanations is essential for achieving reliable and responsible SDD outcomes, moving beyond mere binary classification. By bridging the gap between low-level XAI signals and high-level natural language reasoning, the proposed framework enhances the transparency and trustworthiness of deepfake detection systems. The authors assert that their approach increases "inside accuracy" by over 45% (verified through human evaluation and faithfulness checks) and provides a foundational dataset to support future research in explainable SDD. The work emphasizes that while LLMs offer interpretability, they require the structural support of XAI to remain faithful to the underlying decision process.