Diagnosing and Mitigating Domain Shift in Permission-Based Android Malware Detection

AI systems trained to detect Android malware perform well in controlled settings (92% accuracy) but fail dramatically in real-world deployment, with accuracy dropping to 73% when models trained on one dataset are tested on another—a critical gap that could leave users vulnerable to threats the system was supposed to catch.

The same malware detector can rely on completely different features depending on which dataset it was trained on, meaning a model might identify malware based on one set of permissions in one context but entirely different permissions in another, revealing that these systems are learning dataset-specific quirks rather than genuine malware patterns.

"Noisy" or dataset-specific artifacts in training data are a bigger problem than missing information, suggesting that the solution isn't simply collecting more data, but rather being more careful about what data is used and ensuring it reflects real-world conditions rather than peculiarities of how data was collected.

Explainability tools reveal instability in how AI makes decisions across different contexts, with key features gaining or losing importance unpredictably—this opacity makes it difficult for security teams to trust or understand why the system makes particular malware classifications.

A practical solution exists: training models on only the most common, reliable features across datasets can recover performance (88% accuracy), suggesting that responsible AI deployment requires careful feature engineering and validation across multiple real-world scenarios before deployment.