Abstract

Botnet command-and-control (C2) traffic is among the hardest classes of intrusion to detect: individual bots generate sparse flows indistinguishable from normal host activity. Graph-NOTS introduces the Multi-Scale Persistence Kernel (MSPK), an unsupervised framework that analyzes communication-graph topology via principled multi-scale persistent homology. On the CTU-13 real botnet dataset, MSPK achieves AUC-ROC 0.869 and PR-AUC 0.851 — outperforming the volumetric baseline by +28.7 AUC points — with only ~9.7 ms per-window overhead on commodity hardware, viable for real-time inline NIDS deployment.

Abstract

We investigate whether topological features derived from persistent homology complement hand-crafted ECG features for binary normal/abnormal classification. A fixed-weight kernel fusion (MK-Fusion) combining separate RBF kernels on topological and hand-crafted feature spaces achieves AUC = 0.874 on PTB-XL (21,837 clinical 12-lead ECGs) — the highest among all feature-based classifiers, with statistically significant gains (DeLong's test, p < 0.001). Results indicate persistence landscapes encode complementary geometric information about the cardiac attractor.

Abstract

TritonZ is a remotely operated underwater rover with a manipulator arm for exploration and rescue. Integrates real-time video streaming, advanced navigation, and precise control for challenging aquatic environments.