In short
15,542 programs, roughly half malware. The data was split 50 different ways into training and test sets, and every method used the same 50 splits.
The “Going Beyond API Calls in Dynamic Malware Analysis: A Novel Dataset” (GBAC; Ilić et al., 2024) was used for this research. The GBAC dataset (15,542 samples) was used for its nearly even split of benign and malware samples, 8,217 and 7,325 respectively. The dataset also has a high dimensionality, which directly instantiates the qubit encoding bottleneck: it becomes 262,144 features after hashing.
To reduce the chances or effects of outliers, the dataset was evaluated 50 times using 10 runs of stratified 5-fold cross-validation, with each run shuffled using a different seed from 42 to 51. This resulted in 50 pairs of training and testing splits, which enabled each condition to be compared fold by fold. Each training fold was further split 20/80 into validation and training data for model selection, which meant that testing folds were never used to choose a model.
Stratification kept the class ratios at approximately 53/47 benign to malware, and no resampling or class weighting was used.
Data access
The GBAC dataset is shared under a non-disclosure agreement, so neither the data nor the pipeline code that processes it is published here.