XSec: A Self-Explaining Deep Architecture for Security Applications

Researchers propose a self-explaining deep architecture, XSec, which achieves an average classification accuracy of 97.33% on five security scenarios. During training, the architecture uses a mask-based method to extract informative sub-features and learns prototypes that characterize each class; at test time, it computes similarity scores through a dedicated similarity layer, providing explanations without post-hoc analysis. For a fixed trained model and input, XSec's explanations are deterministic, and compared with approximation- and perturbation-based post-hoc methods, it significantly reduces explanation latency.