Sensitive Region-based Metamorphic Testing Framework using Explainable AI

Explainable AI can identify vulnerability hotspots: The paper demonstrates that Explainable AI (XAI) techniques can pinpoint "sensitive regions" in images where small changes cause AI systems to make different predictions, providing a practical tool for finding weaknesses in image recognition systems before deployment.

Current testing methods miss critical failure points: Most existing tests of deep learning image systems focus on what transformations to apply, but overlook where to apply them. This research shows that targeting sensitive regions is more efficient at detecting misclassifications than untargeted testing approaches.

Small, localized changes can break AI systems: The framework reveals that deep learning models can be surprisingly fragile—even minor alterations to specific image regions can flip the system's predictions, raising concerns about real-world reliability in applications like medical imaging or autonomous vehicles.

Metamorphic testing offers practical quality assurance: By combining metamorphic testing (a method that checks if outputs change appropriately when inputs are transformed) with explainability tools, organizations have a concrete framework to systematically audit and improve their image recognition systems before deployment.

Responsible AI requires understanding failure modes: The research emphasizes that deploying trustworthy AI systems requires not just building models, but actively testing them to understand how and why they fail—knowledge that XAI and targeted testing can provide.