Explainable AI for Zero-Day AttackDetection in IoTNetworks

Loading...
Thumbnail Image

Date

Journal Title

Journal ISSN

Volume Title

Publisher

Unuversity Setif 1 Ferhat Abbas . Faculty of Sciences MAI/1160

Abstract

This thesis presents two deep learning IDS architectures for zero-day attack detection on the Edge-IIoTset dataset (2.2M samples, 15 attack classes, 6 held out for zero-day evaluation). The first, a Cascaded Hierarchical IDS, uses a residual MLP with contrastive and focal loss, Mahalanobis distance in a 64‑D embedding, and score fusion. It achieves 86.64% zero-day recall at 13.00% FPR, F1=0.8654, and 86.86% known‑attack accuracy. The second, a Hybrid CNN‑Based IDS with feature attention and triplet loss, uses Euclidean centroid distances. It achieves 81.02% zero-day recall at 11.09% FPR, F1=0.8418, and 90.94% known‑attack accuracy. Both Frameworktems include multi‑stage explainability (SHAP, Integrated Gradients). Cross‑Frameworktem SHAP reveals the same ten globally important features; Integrated Gradients shows unique anomaly signatures per zero‑day class.

Description

Citation

Endorsement

Review

Supplemented By

Referenced By