ARTFEED — Contemporary Art Intelligence

XGBoost-BalSamp Achieves 95.45% Accuracy in IoT Intrusion Detection

ai-technology · 2026-05-27

Researchers have improved autonomous online intrusion detection for IoT devices, achieving 95.45% accuracy on the UNSW-NB15 benchmark. The study, published on arXiv (2605.26166), first replicates the state-of-the-art AOC-IDS system from IEEE INFOCOM 2024, which uses an Autoencoder with Cluster Repelling Contrastive loss and a Gaussian-based decision module, attaining 89.39% accuracy (close to the published 89.19%). Four key limitations were identified: class imbalance, unreliable pseudo-label generation, limited generalization, and computational overhead. The proposed XGBoost-BalSamp method boosts accuracy by 6.26% over the baseline. Additional deep learning techniques (PseudoFilter, Mix) are also introduced to address these issues, targeting resource-efficient deployment for IoT environments.

Key facts

  • Study published on arXiv with ID 2605.26166.
  • AOC-IDS system from IEEE INFOCOM 2024 is replicated.
  • AOC-IDS uses Autoencoder with CRC loss and Gaussian decision module.
  • Replication achieved 89.39% accuracy on UNSW-NB15.
  • Four limitations identified: class imbalance, unreliable pseudo-labels, limited generalization, computational overhead.
  • XGBoost-BalSamp achieves 95.45% accuracy, a 6.26% improvement.
  • Additional methods include PseudoFilter and Mix.
  • Focus on lightweight architectures for IoT deployment.

Entities

Institutions

  • IEEE INFOCOM
  • arXiv

Sources