Impact of Feature Encoding on Malware Classification Explainability

Feature encoding choices directly affect how well humans can understand AI security decisions: One Hot Encoding (OHE) provided more detailed explanations of why the malware classifier made specific predictions compared to Label Encoding (LE), even though LE performed slightly better at classification accuracy.

Better explainability can outweigh small performance trade-offs in security applications: The researchers found that OHE's superior explanations compensated for its marginal performance loss, suggesting that in malware detection—where understanding why a threat was flagged matters for security teams—explainability may be more valuable than marginal accuracy gains.

Explainability improvements can also reduce analyst workload: Using OHE resulted in smaller explanation files and shorter analysis times for human experts reviewing the AI's decisions, making the security review process more efficient alongside being more transparent.

Technical implementation details have real consequences for responsible AI: This research demonstrates that seemingly low-level decisions about how to prepare data for AI systems significantly impact whether security professionals can trust, verify, and act on the AI's recommendations—a critical requirement for responsible deployment in high-stakes domains.