Virginia Tech
Tenko++: Next-Gen Lightweight Intrusion Detection with Context-Aware Anomaly Scoring
Abstract
dc:description.abstractAs modern networks face increasingly sophisticated cyber threats, achieving lightweight yet robust intrusion detection remains a critical challenge, especially in dynamic and resource-constrained environments. Although frameworks such as Kitsune demonstrate the feasibility of online unsupervised anomaly detection using autoencoder ensembles, they remain sensitive to benign traffic variability, static threshold selection, and the absence of historical context in packet-level decision-making. This thesis presents Tenko++, a lightweight, context-aware intrusion detection framework that extends Kitsune-style reconstruction-based anomaly detection with skip-connected autoencoders, node-level behavioral scoring, adaptive thresholding, and Zero Trust-aligned trust evaluation. Tenko++ converts volatile packet-level reconstruction errors into persistent node-level risk signals, allowing anomaly decisions to incorporate historical behavior rather than relying only on instantaneous RMSE values. Tenko++ is evaluated across nine attack scenarios under benign-only training and realistic thresholding assumptions. The results show improved detection stability over Kitsune, particularly for sustained behavioral attacks such as Mirai Botnet and SSDP Flood. Under realistic Mirai thresholding, Kitsune achieves a TPR of 0.883 and F1 score of 0.937, while Tenko++ achieves approximately 0.998 TPR and 0.998 F1 with near-zero false positives. Additional experiments evaluate performance under distribution shift, memory constraints, poisoning, smoothing, distance metric selection, ablation settings, and real-hardware deployment. Hardware evaluation shows that Tenko++ remains practical for edge deployment, requiring approximately 1.168 ms per packet and 0.169 MB of memory on Raspberry Pi 3B. These results demonstrate that Tenko++ preserves the lightweight nature of Kitsune while improving stability through context-aware, persistence-driven anomaly interpretation. Overall, the thesis shows that effective unsupervised intrusion detection in deployment-constrained environments depends not only on detecting high anomaly scores, but on interpreting those scores through node history, adaptive thresholds, and operational context.
Degree
thesis:*- Name thesis:degree_name
- Master of Science
- Level thesis:degree_level
- masters
- Discipline thesis:degree_discipline
- Computer Engineering
- Department dc:contributor.department
- Electrical and Computer Engineering
- Grantor dc:publisher
- Virginia Tech
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Dave, Daksh Rajesh
- Chair dc:contributor.committeechair
-
- Burger, Eric William
- Committee members dc:contributor.committeemember
-
- Tripathi, Nishithkumar Dhananjay
- Midkiff, Scott F.
Subjects
dc:subject × 7Rights
dc:rights- Statement dc:rights
-
- In Copyright
- Licence dc:rights.uri
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Dc Identifier Other
- vt_gsexam:46874