XAI-Grounded Explanation Generation for Speech Deepfake Detection with Training-Free Multimodal Large Language Models
Researchers have developed a novel training-free framework to enhance speech deepfake detection systems by generating explanations grounded in XAI evidence and multimodal large language models. This approach aims to overcome the limitations of traditional XAI methods, which produce low-level attribution signals, and LLM-based generation, which often lacks specificity due to limited grounded explanation datasets for speech deepfake detection. By integrating XAI with multimodal LLMs and constructing a new grounded explanation dataset from the PartialSpoof dataset, the framework has demonstrated an increase in accuracy of over 45% through human evaluation and faithfulness checks. AI
IMPACT Improves trustworthiness of AI systems by providing understandable explanations for deepfake detection.