Skip to main navigation Skip to search Skip to main content

HTTP2DoS: Detecting low-and-slow HTTP/2 DoS attacks with complex transfer-entropy graph-temporal learning

Research output: Contribution to journal › Article › peer-review

Abstract

Application-layer denial-of-service (DoS) attacks increasingly threaten web applications and APIs by exhausting server resources without high traffic volumes. The design of HTTP/2 facilitates low-and-slow (LnS) attacks that mimic legitimate client behavior and evade traditional detection mechanisms. This paper presents HTTP2DoS, a graph-temporal framework for detecting LnS HTTP/2 DoS attacks against APIs. HTTP2DoS models decoded HTTP/2 event activity as a multivariate time series and constructs a directed dependency graph over protocol-aware feature variables using transfer entropy. The resulting graph is integrated into a forecasting model that captures short-term frame activity, long-range timing dependencies, and asymmetric relationships among HTTP/2 behaviors. To model coordinated pacing patterns, the teacher model uses complex-valued hidden representations and resonance-weighted graph propagation, while a lightweight student approximates the teacher through graph sketch diffusion and geometric knowledge distillation, reducing inference cost while preserving detection performance. We evaluate HTTP2DoS on a realistic virtual banking testbed, where autonomous multi-agent personas generate benign API traffic and controlled simulations produce diverse LnS HTTP/2 attack scenarios. HTTP2DoS is evaluated under two scoring modes, a retrospective mode for out-of-band analysis of recorded traffic and a real-time mode that scores each event using only preceding observations. HTTP2DoS improves segment-adjusted F1 by 7.1% over the strongest state-of-the-art multivariate time-series anomaly detection baseline and improves event-sequence classification accuracy by 13.2% over the HTTP/2 event-sequence detection baseline.

Original languageEnglish
Article number105198
JournalComputers and Security
Volume172
DOIs
StatePublished - Jan 2027

Keywords

  • Anomaly detection
  • Graph neural network
  • HTTP/2
  • Knowledge distillation
  • Low and slow DoS attacks
  • Multivariate time series

ASJC Scopus subject areas

  • General Computer Science
  • Law

Fingerprint

Dive into the research topics of 'HTTP2DoS: Detecting low-and-slow HTTP/2 DoS attacks with complex transfer-entropy graph-temporal learning'. Together they form a unique fingerprint.

Cite this