Cyber-Attack Variant Prediction for Long-Term ML Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing cyber-attack detection systems struggle to detect new variants of known threats due to evolving attack patterns, leading to ineffective ML models over time, especially in dynamic environments.

Innovation Solution

A cyber-polymorphism system using a Siamese neural network trained with predicted network metadata generated by a conditional generative adversarial network (CT-GAN) to identify and generate new variants of known cyber-attacks, enhancing long-term detection capabilities.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If ML models are trained on known attack patterns, then detection accuracy for known threats is improved, but the models become ineffective against new variants over time

Engineering Contradiction:
Improvedetection accuracyVSAvoidmodel effectiveness over time
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The system performs preliminary actions by generating predicted variants of known cyber-attacks before actual new variants appear. A conditional GAN creates synthetic attack variants based on known patterns, and these predicted variants are used to pre-train the ML model, enabling it to anticipate and detect new variants before they emerge in the wild.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system makes the ML model dynamic by continuously updating its training data with newly predicted variants. Instead of using static training sets, the system generates new predicted variants periodically and re-trains the model, allowing it to adapt to evolving attack patterns over time without requiring manual intervention.

Inventive Principle:
Principle #15Dynamics

2Adaptability or versatility

If ML models are frequently retrained to keep up with evolving threats, then detection capability is improved, but computational resources and time are consumed

Engineering Contradiction:
Improvedetection capabilityVSAvoidretraining time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The conditional GAN performs preliminary generation of attack variants that anticipate future threats, allowing the ML model to be trained on predicted variants rather than waiting for actual new variants to appear. This preliminary action reduces the frequency and urgency of retraining operations.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system creates synthetic copies of attack patterns through the conditional GAN, which generates predicted variants that mimic the structure and characteristics of real attacks. These synthetic copies serve as training data, eliminating the need to collect and process large amounts of real attack data for each retraining cycle.

Inventive Principle:
Principle #26Copying

3Ease of operation

If signature-based detection is used, then known attacks are easily detected, but new variants with different patterns cannot be detected

Engineering Contradiction:
Improvedetection simplicityVSAvoidvariant detection capability
Core Design Contradiction:
Ease of operationVSAdaptability or versatility

Solution Approach 1:

The system transforms the static signature-based approach into a dynamic system that continuously generates new signatures through the conditional GAN. Instead of relying on fixed signatures, the model learns to recognize patterns and can adapt to new variants by generating updated detection signatures from predicted attack variants.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The conditional GAN changes the parameters of attack patterns by generating predicted variants with modified characteristics while maintaining the core attack structure. This allows the system to detect variants that differ from known patterns in terms of timing, sequence, or specific parameters, while still recognizing them as members of the same attack family.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4704376A1Long-term cyber attack variant prediction
Publication Date: 2026.03.04 BRITISH TELECOM PLC
  • EP4704376A1 patent drawingFigure 1
  • EP4704376A1 patent drawingFigure 2a
  • EP4704376A1 patent drawingFigure 2b

AI summary

A computer-implemented method of generating training data for a cyber-attack detection machine learning, 'ML', model, the computer-implemented method comprising: obtaining a network metadata feature set characteristic of each of one or more known categories of cyber-attack; and generating predicted network metadata representing a plurality of predicted variants of each of the one or more known categories of cyber-attack, based on the respective network metadata feature set.