Host Identifier Classification in Encrypted Traffic Using Flow Embeddings

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing network providers struggle to correlate encrypted network traffic with upper-layer behaviors, such as identifying which web page a network flow corresponds to, especially when multiple pages are loaded simultaneously, as encrypted traffic obscures the distinction between core and support domains.

Innovation Solution

A network device employs unsupervised learning to create high-dimensional vectors for host identifiers, followed by supervised learning using a labeled ground truth to classify these identifiers into groups, utilizing techniques like natural language processing and word embeddings, enabling the differentiation between core domains and support domains within encrypted traffic.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If network traffic is encrypted to protect user privacy and security, then security and privacy are improved, but the ability to correlate network flows with web page behaviors is lost

Engineering Contradiction:
Improvesecurity and privacyVSAvoidloss of correlation between network flows and web page behaviors
Core Design Contradiction:
ReliabilityVSLoss of information

Solution Approach 1:

The patent introduces flow logs as an intermediary that captures metadata about network flows (source/destination IPs, ports, timestamps, durations) without requiring decryption of the actual traffic content. This mediator enables analysis and correlation of network behaviors with web page activities while maintaining encryption of the actual communication content, thus resolving the contradiction between security and analyzability.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Productivity

If multiple network flows are opened to download resources for web pages, then web page functionality is improved, but the ability to distinguish core domains from support domains is degraded

Engineering Contradiction:
Improveweb page functionalityVSAvoidcomplexity of distinguishing core and support domains
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent applies preliminary action by using unsupervised learning to pre-process and embed flow log data into high-dimensional vectors before classification. This preliminary transformation organizes the complex multi-flow data structure, making subsequent supervised classification of core versus support domains more effective and less complex.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent transforms the original flow log data into high-dimensional vector space through unsupervised learning embeddings. This dimensionality change allows the system to capture complex relationships between multiple network flows and their corresponding web page resources, enabling effective distinction between core and support domains despite the multiplicity of flows.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Measurement precision

If unsupervised learning is used to create high-dimensional vectors for host identifiers, then classification precision is improved, but computational complexity increases

Engineering Contradiction:
Improveclassification precisionVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies preliminary action by using unsupervised learning to pre-process and embed flow log data into high-dimensional vectors before classification. This preliminary transformation organizes the complex multi-flow data structure, making subsequent supervised classification of core versus support domains more effective and less complex.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent transforms the original flow log data into high-dimensional vector space through unsupervised learning embeddings. This dimensionality change allows the system to capture complex relationships between multiple network flows and their corresponding web page resources, enabling effective distinction between core and support domains despite the multiplicity of flows.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Data Source

PatentEP4111380B1Network device and method for host identifier classification
Publication Date: 2026.03.25 HUAWEI TECH CO LTD
  • EP4111380B1 patent drawingFigure 1
  • EP4111380B1 patent drawingFigure 2
  • EP4111380B1 patent drawingFigure 3

AI summary

The present disclosure relates to the field of computer networks. More specifically, a solution for machine learning-based classification of host identifiers in encrypted network traffic is provided. The classification can, in particular, include natural language processing capabilities. The present disclosure provides a network device (100) for host identifier classification. The network device (100) is configured to obtain a sequence (101) of host identifiers (102), each host identifier (102) corresponding to a flow of encrypted network traffic (103), apply an unsupervised learning technique (104) to the sequence (101) of host identifiers (102) to learn a vector (105) of a high-dimensional space for each host identifier (102) in the sequence (101), obtain a labelled ground truth (106) comprising labels (107) corresponding to a host identifier (102), and apply a supervised learning technique (108) to each vector (105), based on the labelled ground truth (106), to classify the corresponding host identifier (102).