Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Introducing Adaptive Continuous Adversarial Training (ACAT) to Enhance ML Robustness

Mohamed elShehaby, Aditya Kotha, Ashraf Matrawy

TL;DR

ACAT addresses the scarcity of labeled adversarial data and the challenge of concept drift in cybersecurity by continuously injecting real-world detected adversarial samples into ongoing model training, using Elastic Weight Consolidation to mitigate forgetting. The method is validated on a problem-space SPAM filtering task with TextFooler-based perturbations, demonstrating robustness gains (accuracy rising from ~69% to ~88% after three retraining sessions) and significantly faster inference than conventional two-model approaches. These findings suggest that continual, adaptive adversarial training can provide practical, scalable defenses against evolving attacks in security-sensitive domains, reducing retraining overhead while maintaining performance on clean data.

Abstract

Adversarial training enhances the robustness of Machine Learning (ML) models against adversarial attacks. However, obtaining labeled training and adversarial training data in network/cybersecurity domains is challenging and costly. Therefore, this letter introduces Adaptive Continuous Adversarial Training (ACAT), a method that integrates adversarial training samples into the model during continuous learning sessions using real-world detected adversarial data. Experimental results with a SPAM detection dataset demonstrate that ACAT reduces the time required for adversarial sample detection compared to traditional processes. Moreover, the accuracy of the under-attack ML-based SPAM filter increased from 69% to over 88% after just three retraining sessions.

Introducing Adaptive Continuous Adversarial Training (ACAT) to Enhance ML Robustness

TL;DR

ACAT addresses the scarcity of labeled adversarial data and the challenge of concept drift in cybersecurity by continuously injecting real-world detected adversarial samples into ongoing model training, using Elastic Weight Consolidation to mitigate forgetting. The method is validated on a problem-space SPAM filtering task with TextFooler-based perturbations, demonstrating robustness gains (accuracy rising from ~69% to ~88% after three retraining sessions) and significantly faster inference than conventional two-model approaches. These findings suggest that continual, adaptive adversarial training can provide practical, scalable defenses against evolving attacks in security-sensitive domains, reducing retraining overhead while maintaining performance on clean data.

Abstract

Adversarial training enhances the robustness of Machine Learning (ML) models against adversarial attacks. However, obtaining labeled training and adversarial training data in network/cybersecurity domains is challenging and costly. Therefore, this letter introduces Adaptive Continuous Adversarial Training (ACAT), a method that integrates adversarial training samples into the model during continuous learning sessions using real-world detected adversarial data. Experimental results with a SPAM detection dataset demonstrate that ACAT reduces the time required for adversarial sample detection compared to traditional processes. Moreover, the accuracy of the under-attack ML-based SPAM filter increased from 69% to over 88% after just three retraining sessions.
Paper Structure (17 sections, 6 figures, 2 tables)

This paper contains 17 sections, 6 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Adversarial Evasion Attacks and Adversarial Training
  4. Continuous Machine Learning
  5. Our Proposed Adaptive Continuous Adversarial Training (ACAT)
  6. Experimental Evaluation using Problem-Space Adversarial Samples
  7. Related work on ML in SPAM Filters
  8. Problem-space adversarial attacks against SPAM Filters
  9. Dataset and Adversarial Samples in our experiments
  10. Detailed pre-processing for the SPAM problem
  11. Results and Discussion
  12. Hybrid Bi-LSTM SPAM filter performance
  13. Adversarial sample detector performance
  14. ACAT vs Conventional Model
  15. Fine Tuning vs EWC in Adversarial Continuous Training
  16. ...and 2 more sections

Figures (6)

  • Figure 1: Conventional Detecting Approach
  • Figure 2: The Proposed Adaptive Continuous Adversarial Training (ACAT)
  • Figure 3: Pre-processing of the Proposed approach, Adaptive Continuous Adversarial Training (ACAT)
  • Figure 4: Prediction time of ACAT vs a Conventional Approach
  • Figure 5: Accuracy of Fine Tuning vs EWC on the original training set (without adversarial perturbations) during Adversarial Continuous Training. The solid black line represents the 10 training epochs before deployment, while the dotted lines represent the accuracy after each adversarial training session.
  • ...and 1 more figures