SHAP Dilutes Malware Feature Credit by 1/m and Can Reverse the Sign of an Unused Feature

arXiv 2609.04626·medium signalXBlueskyLinkedInCopy link

arXiv 2609.04626 argues SHAP's formal properties, local accuracy, missingness and consistency, are insufficient for malware interpretation because SHAP explains a chosen feature-coalition game, not behavior in the data, and that game is only fixed once the analyst picks the feature players, the missing-feature rule, the background distribution and the simplified input mapping. In static Portable Executable feature spaces, byte histograms, byte-entropy, strings, headers, sections, imports and data-directories are jointly shaped by file structure, packing and compiler behavior rather than being independent. The paper proves conditional SHAP dilutes a model's feature credit by 1/m across m-1 redundant features, attributes importance to features the model never uses, and can reverse the sign of an unused feature's attribution when the distribution shifts.