abamidele/Latent-Stability-Analysis-of-Malware-Representations-Under-Feature-Space-Perturbations

# Latent Stability Analysis of Malware Representations Under Feature-Space Perturbations This repository provides the experimental code for latent-stability analysis of malware feature representations under controlled feature-space perturbations. The pipeline evaluates full EMBER/BODMAS-style PE feature vectors, PCA-compressed features, beta/denoising VAE latent representations, Mandelbrot-inspired escape-time descriptors, Latent Escape Divergence (LED), and PINNFlow-based residual, velocity, risk, and gradient-shift diagnostics. The goal is not to replace high-performing clean-sample malware classifiers, but to provide additional robustness and representation-stability diagnostics that help explain how malware feature representations change under perturbation. The code supports clean classification ablation, perturbation robustness evaluation, bootstrap confidence intervals, McNemar paired tests, boundary-complexity estimates, and optional externally supplied clean/perturbed PE feature-pair evaluation. This repository works on extracted feature vectors and does not generate or modify malware binaries.