Unveiling Hidden DNN Defects with Decision-Based Metamorphic Testing

Current AI testing methods miss significant defects: Existing metamorphic testing frameworks fail to detect a "considerable number" of hidden defects in deep neural networks (DNNs) because they only check if outputs remain consistent after image mutations, missing cases where the AI reaches correct answers through flawed reasoning.

AI systems can make right predictions for wrong reasons: DNNs may rely on erroneous decision-making processes—focusing on incorrect image regions—yet still produce correct outputs by chance. These hidden defects would likely cause failures when the AI encounters diverse real-world images, creating safety risks that current testing overlooks.

Explainable AI tools can reveal hidden problems: The research uses explainable AI (XAI) techniques to identify which visual concepts and image regions the DNN actually uses to make decisions, enabling detection of cases where the AI's reasoning process is flawed even when outputs appear correct.

Testing needs to examine how AI decides, not just what it decides: The paper demonstrates that responsible AI testing must go beyond verifying consistent outputs and instead validate that AI systems are making decisions based on appropriate reasoning—a critical distinction for deploying trustworthy AI systems in real-world applications.