Activation-Deactivation: A General Framework for Robust Post-hoc Explainable AI

Addresses a fundamental problem with current explanation methods: The paper identifies that popular perturbation-based explainability techniques (which explain AI decisions by testing what happens when inputs are altered) rely on unrealistic, out-of-distribution modifications, potentially producing misleading explanations about how models actually work.

Proposes a practical alternative approach: Rather than artificially modifying input data, the Activation-Deactivation method "switches off" internal parts of the AI model itself to understand which components drive specific predictions, avoiding the need to create fake or unrealistic data scenarios.

Designed for easy real-world implementation: ConvAD, the implementation for image-recognition AI systems, works as a "drop-in mechanism" that can be added to already-trained models without retraining, making it practical for organizations to adopt without significant computational costs or disruption.

Demonstrates improved reliability and generalization: The paper shows ConvAD produces explanations that are both more robust (consistent and reliable) and more transferable (work better across different models and datasets) compared to existing state-of-the-art methods, suggesting more trustworthy insights into model behavior.

Validated across diverse conditions: The evaluation spans multiple neural network architectures, datasets, and perturbation strategies, indicating the findings are broadly applicable rather than limited to specific scenarios.