DeepFaith: A Domain-Free and Model-Agnostic Unified Framework for Highly Faithful Explanations

DeepFaith addresses a fundamental problem in AI transparency: existing explanation methods lack objective evaluation standards because there's no agreed-upon "ground truth" for what makes an explanation good, making it difficult to know which explanation method to trust.

The framework achieves theoretical optimality across multiple faithfulness measures: by unifying several widely-used faithfulness metrics into a single objective, DeepFaith provides explanations that simultaneously perform well on 10 different evaluation metrics, offering a more robust measure of trustworthiness than methods optimized for single metrics.

The system works universally across different AI models and domains: tested on 6 different models and 6 datasets, DeepFaith outperformed all baseline methods without needing to be retrained for each new model or application area, suggesting it could be practically deployed across diverse AI systems.

Explanations can be generated efficiently without accessing the original model: once trained, DeepFaith generates explanations through a single computational pass, making it practical for real-world deployment where you may not have direct access to proprietary AI systems.

The approach combines multiple existing explanation methods to improve reliability: rather than replacing current XAI techniques, DeepFaith learns from multiple explanation methods simultaneously, filtering and combining their insights to produce more faithful explanations than any single method alone.