Cloud-based XAI Services for Assessing Open Repository Models Under Adversarial Attacks

XAI tools need better infrastructure: Current explainable AI tools aren't designed to work as scalable services, so the authors built a cloud-based framework to automate XAI assessments—highlighting a practical gap in how organizations can deploy and manage these transparency tools at scale.

Adversarial attacks undermine both predictions and explanations: The research tested five quality attributes of AI models and found that adversarial attacks don't just fool model predictions; they also compromise the reliability of explanations themselves through "explanation deviation" and reduced "explanation resilience," meaning users can't trust the reasoning behind predictions when models are under attack.

Assessing AI robustness requires testing many combinations: The framework evaluated over a hundred different scenario combinations across computer vision and tabular data, demonstrating that responsible AI validation is complex and labor-intensive—suggesting organizations need systematic, automated approaches rather than ad-hoc testing.

Multiple quality dimensions matter for responsible deployment: Beyond accuracy, the research identified five critical quality attributes (computational cost, performance, robustness, explanation deviation, and explanation resilience) that should be evaluated before integrating open-source AI models into real services, providing a more comprehensive view of what "safe to deploy" means.

Open-source models require rigorous validation before use: The paper emphasizes that the opacity of AI models necessitates thorough validation and evaluation before integration into services, particularly when using publicly available models where you may not control the training process.