PaperlayerView exact revision on arXiv

Source-linked AI summary

A GAN-Based Framework for Robust DDoS Attack Detection

Makram Chehayeb, Walid Fahs, Amina Rizk, Rida Khatoun, Omran Berjawi

arXiv:2609.18281v1cs.LGcs.CR

TL;DR

DDoS detectors can lose effectiveness when attackers craft adversarial traffic that evades learned decision boundaries. The paper augments CICDDoS2019 training data with WGAN-GP-generated adversarial flows and evaluates Random Forest, Deep Neural Ensemble, and Transformer models under normal, adversarial, and simulated deployment conditions. The reported results indicate improved detection resilience, particularly for unseen adversarial traffic, while highlighting practical deployment considerations.

  • Problem

    Modern DDoS attacks can evade traditional and machine-learning defenses, threatening the availability and consistency of online services.

  • Method

    The framework combines WGAN-GP-generated adversarial flows with benign and malicious CICDDoS2019 traffic to train Random Forest, Deep Neural Ensemble, and Transformer detectors.

  • Results

    The reported experiments show that adversarial augmentation strengthens DDoS detectors against feature-space evasion, with the Transformer showing consistent high performance across evaluated phases.

  • Takeaways & Limitations

    The framework supports more resilient and adaptive DDoS detection by combining adversarial data augmentation with modern machine-learning models and simulated operational evaluation.

Abstract

from arXiv · show

The availability and consistency of online services remain vulnerable due to Distributed Denial of Service (DDoS) attacks. These attacks are evolving by adopting more complex strategies to evade traditional network security systems. Despite the effectiveness of machine learning models in detecting DDoS traffic, targeted adversarial attacks can degrade their classification accuracy. This work proposes a robust detection framework that integrates generative adversarial modelling with advanced machine learning models. We trained Random Forests, Deep Neural Ensembles, and Transformer-based models using the CICDDoS2019 dataset to establish the frameworks baseline performance. To enhance the models defensive capacity, we generated synthetic adversarial flows that simulate potential evasion attempts and adversarial traffic using a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP). Then, we combined the generated traffic with benign and malicious traffic to construct hybrid datasets to train the models to learn more generalizable decision boundaries. The experimental results indicate that the proposed methodology significantly enhances detection accuracy and resilience, especially against unseen adversarial traffic. We also tested the designed framework using real-world generated traffic, which demonstrates its capability in practical settings. The scalable and efficient solution against adversarial DDoS attacks, introduced in this work, paves the way towards more resilient and adaptive network defense systems that combine generative adversarial augmentation with recent advances in learning models.

I. INTRODUCTION

Modern DDoS campaigns exploit expanding cloud, IoT, and botnet ecosystems to bypass legacy and AI-based defenses. This work proposes WGAN-GP adversarial augmentation and evaluates detection models for robustness, operational behavior, and deployment feasibility.

  • I. INTRODUCTION: DDoS campaigns are becoming larger and more complex as cloud services, IoT connectivity, and botnet availability expand, threatening uptime-dependent services.The cited examples include remote medical telesurveillance and smart home health monitoring.
  • I. INTRODUCTION: Adversarial attacks can alter malicious traffic features so classifiers interpret the traffic as benign, exposing a key weakness of modern detection systems.Traditional IDS also generalize poorly to new attack variants.
  • I. INTRODUCTION: WGAN-GP generates diverse synthetic adversarial DDoS flows that are combined with benign and malicious traffic to harden detectors against evasion.The augmented data are intended to improve generalization while preserving realistic attack characteristics.
  • I. INTRODUCTION: The study benchmarks clean-trained Random Forest, Deep Neural Ensemble, and Transformer architectures under feature-space evasion and evaluates their recovery after adversarial augmentation.The benchmark specifically considers Adv-5 and Adv-9 evasion conditions.
  • I. INTRODUCTION: The framework also examines protocol-aware feature realizability, three-phase live deployment behavior, and latency, throughput, and memory costs for network-edge operation.The operational phases are Before/Normal, During/Attack Peak, and After/Recovery.

II. LITERATURE REVIEW

DDoS detection research has progressed toward deep learning and GAN-based adversarial defense, but important deployment gaps remain. Prior work commonly emphasizes static benchmark performance while underexamining operational shifts, feature realizability, and resource overhead.

  • II. LITERATURE REVIEW: Classical classifiers can exceed 99% accuracy on clean CAIDA 2007 and CICDDoS2019 benchmarks but degrade severely under subtle feature perturbations.The cited models include Decision Trees, Naïve Bayes, SVMs, and Random Forests.
  • II. LITERATURE REVIEW: Reported systems include LSTM-based GAN augmentation with up to 100% accuracy and approximately 0.99 F1, plus DBN-LSTM detection with 96.55% accuracy, 98.53% recall, and 97.47% F1.The DBN-LSTM system maintained detection above 91% under FGSM testing, while the GAN-based system retained 91.7%–[incomplete range] under evasion.
  • II. LITERATURE REVIEW: GAN-based adversarial training can improve resilience against adaptive cyberattacks, but existing evaluations remain mainly offline and static rather than real-time.The literature also lacks sufficient attention to practical integration into large-scale ISP-level defenses.
  • II. LITERATURE REVIEW: Existing defenses span ISP filtering, scrubbing centers, cloud platforms, on-premise appliances, hybrid deployments, and application-level controls.These layers coordinate filtering and threat intelligence across network operators and protected organizations.
  • II. LITERATURE REVIEW: Prior GAN augmentation methods often modify feature vectors directly without validating whether generated values satisfy networking logic such as packet-rate and byte-count consistency.This creates a feature-realizability gap between mathematical perturbations and protocol behavior.

III. METHODOLOGY

The methodology models DDoS traffic through an ISP queuing network and applies an AI detector at the last ISP router. The detector extracts flow features, predicts maliciousness, and discards flows exceeding a tunable threshold.

  • A. DDoS attacks model: The framework models botnet traffic traversing ISP routers as a directed multi-stage queuing network with finite link capacities, delays, and router buffers.Each router is represented as a finite-buffer FIFO queue.
  • A. DDoS attacks model: Legitimate arrivals follow a Poisson process, whereas attack traffic uses an MMPP with state-dependent rates to represent burstiness.The attack process is parameterized by a transition matrix Q and m intensity states.
  • A. DDoS attacks model: Router service is exponential with rate µ_n = C_n/L_pkt, capped by link capacity, while finite buffers produce packet drops under overflow.Queue behavior follows an M/M/1/K model with steady-state probabilities π_n,j.
  • A. DDoS attacks model: At the last ISP router, the system extracts packet-flow features and classifies traffic using RF, DNE, or TF models.Features include inter-arrival time, packet size and entropy, IP reputation, flow duration, and connection rate.
  • A. DDoS attacks model: Flows with p_malicious > θ are discarded; otherwise, they are forwarded, with detection delay, false-positive rate, and false-negative rate tracked.The detector is periodically retrained to adapt to evolving attack patterns.

B. CICDDoS2019 Dataset

The study uses CICDDoS2019, a flow-based dataset containing benign traffic and multiple DDoS attack types. Preprocessing standardizes the data and retains the 20 most discriminative features.

  • B. CICDDoS2019 Dataset: CICDDoS2019 contains over 80 CICFlowMeter-derived flow features spanning benign traffic and multiple DDoS attack families.Listed attacks include UDP, HTTP, SYN, DNS, MSSQL, LDAP, NTP, NetBIOS, SSDP, UDP-Lag, and WebDDoS.
  • B. CICDDoS2019 Dataset: The dataset is imbalanced, with more attack instances than benign instances.
  • B. CICDDoS2019 Dataset: Preprocessing removes non-informative columns, imputes missing values, standardizes features, and selects the top 20 features using the ANOVA F-test.SelectKBest is used for feature selection in all experiments.

C. Adversarial Datasets

The adversarial-dataset method uses WGAN-GP to model attack-flow distributions and generate synthetic traffic. Its objective combines Wasserstein distribution matching with a gradient penalty enforcing a Lipschitz constraint.

  • C. Adversarial Datasets: WGAN-GP uses a Generator and Critic in a minimax framework to generate synthetic adversarial attack traffic.The real attack distribution comes from CICDDoS2019.
  • C. Adversarial Datasets: The standard WGAN objective measures the difference between real and generated distributions using the Wasserstein distance.The critic is optimized over the set of 1-Lipschitz functions.
  • C. Adversarial Datasets: WGAN-GP adds a gradient penalty to enforce the critic’s Lipschitz constraint during training.The penalty is evaluated on random interpolations between real and generated samples, with λ set to 10.

1) Architecture Specifications: •

The generator maps 100-dimensional Gaussian noise to 20-dimensional synthetic attack feature vectors. The critic maps feature vectors to scalar realness scores and is updated more frequently than the generator.

  • 1) Architecture Specifications: •: The generator uses Dense layers of sizes 128, 256, and 20 with ReLU activations internally and Tanh output to produce synthetic attack vectors.Its input is 100-dimensional noise sampled from N(0, 1).
  • 1) Architecture Specifications: •: The critic uses Dense layers of sizes 256, 128, and 1, with LeakyReLU activations and a linear scalar realness output.The LeakyReLU slope is 0.2.
  • 1) Architecture Specifications: •: Training runs for 100 epochs with batch size 256 using Adam, while the critic receives four updates per generator update.The learning rate is 2×10^-4, with β1 = 0.5 and β2 = 0.999.

2) Adversarial Dataset Generation:

The framework generates synthetic attack flows with WGAN-GP and perturbs selected features to create adversarial datasets for training and robustness evaluation.

  • WGAN-GP-generated attack flows are perturbed in targeted features to mimic benign traffic while retaining malicious functionality.
  • The perturbation procedure uses Fmod to identify modified features and samples benign-distribution values for replacement.
  • Hybrid training datasets combine CICDDoS2019 with adversarial samples modifying 2, 4, or 6 randomly selected features.
  • Adv-5 and Adv-9 test sets perturb 5 or 9 most discriminative features ranked by ANOVA F-score.

D. Detection Models

The study compares Random Forest, Deep Ensemble Neural Network, and Transformer detectors, including a Transformer architecture designed to capture global feature dependencies.

  • The Random Forest baseline uses 100 unrestricted-depth trees with balanced class weighting to address dataset imbalance.
  • The Deep Ensemble averages predictions from five independently trained feed-forward neural networks with identical dropout architectures.
  • The framework evaluates Random Forest, Deep Ensemble Neural Network, and Transformer models, with the Transformer intended to capture global dependencies among traffic features.
  • The Transformer embeds 20-dimensional inputs into 32 dimensions and uses four attention heads, normalization, residual links, and a classification head.

IV. EXPERIMENTAL SETUP AND RESULTS

Clean-trained detectors perform poorly on adversarial traffic, whereas WGAN-GP augmentation improves robustness, particularly for the Transformer model.

  • A. Phase 1: Baseline Model Performance: The baseline evaluation used clean CICDDoS2019 training data and tested models on clean, Adv-5, and Adv-9 traffic, prioritizing Recall to avoid missed attacks.
  • A. Phase 1: Baseline Model Performance: Baseline performance dropped sharply on adversarial traffic, with Adv-9 Recall falling to 0.1036 for RF and 0.0465 for DNE, while TF reached 0.2045.
  • B. Phase 2: Robustness via Adversarial Augmentation: The Transformer reached 0.8002 Recall on the hardest Adv-9 evaluation after adversarial augmentation, while the DNE remained the weakest model there.
  • B. Phase 2: Robustness via Adversarial Augmentation: Adversarial augmentation strengthened detectors against feature-space evasion, with the Transformer showing consistently high performance across both evaluation phases.

C. Real-Time Traffic Evaluation

Real-time simulation across normal, attack-peak, and recovery traffic exposed different operational behaviors, with the Transformer providing the most balanced deployment performance.

  • C. Real-Time Traffic Evaluation: The simulation used three unlabeled traffic segments representing normal operation, attack peak, and recovery, with inference performed flow-by-flow without additional training.
  • C. Real-Time Traffic Evaluation: The RF produced no normal-operation false alarms but detected almost none of the attack peak, reaching 0.01 Recall.
  • C. Real-Time Traffic Evaluation: The DNE labeled all traffic as malicious across normal, attack, and recovery phases, producing a 100% false-positive rate during normal operation.
  • C. Real-Time Traffic Evaluation: The Transformer achieved perfect normal-operation accuracy, 0.69 attack-phase Recall, and detection of malicious traffic during recovery.
  • D. Discussion: These results support adversarial augmentation as an efficient strategy for improving generalization and resilience, especially for RF and TF models.

V. CONCLUSION AND FUTURE WORK

The study finds that DDoS detection models face practical trade-offs and reduced adaptability in real-world adversarial settings. Future work targets packet-level realism, broader cross-dataset validation, and online adaptation.

  • RF performs well on normal traffic but fails during active attacks, while Deep Ensembles achieve high recall at the cost of excessive false positives.The Transformer model achieves high recall and sensitivity but requires further tuning to improve precision.
  • Future work will test packet-level evasion tools to assess whether generated attacks are realizable under protocol constraints in physical testbeds.
  • The framework will be validated across additional datasets and extended with online learning for real-time updates under dynamic zero-day attack patterns.
Loading 2609.18281v1…