Aymen Bouferroum

Most Wi-Fi sensing models are trained on simulated data built by adding noise to recorded channel estimates. We tested that assumption on six commodity receivers and found it breaks: the receiver's automatic gain control compresses the signal in ways no additive noise can reproduce. Our answer is MQTC, a measurement-calibrated model combining quantile mapping, temporal filtering, and copula-based reordering, which cuts amplitude error 8-fold and closes 89% of the fidelity gap. Classifiers trained on MQTC data recover 93% of real-world jamming-detection performance, while noise-trained ones remain near random.

@misc{bouferroum2026csisimulationadditivenoise,
      title={CSI Simulation: Why Additive Noise Fails and How to Fix It},
      author={Aymen Bouferroum and Ildi Alla and Vincent Lenders and Valeria Loscri},
      year={2026},
      eprint={2607.01882},
      archivePrefix={arXiv},
      primaryClass={cs.NI},
      url={https://arxiv.org/abs/2607.01882},
}

A jamming attack can silence an entire wireless factory floor. CITADEL detects and classifies jamming using only Channel State Information, a signal that commodity IIoT devices already produce for free, and it recognizes attacks it has never been trained on. Across 6 known and 15 zero-day attack scenarios it reaches 100% and 97.1% detection at a 0.4% false-positive rate, resists adversarial evasion, and runs end-to-end in 14.2 ms on an edge GPU, outperforming eight baselines.

@misc{bouferroum2026citadelcsibasedjammingdetection,
      title={CITADEL: CSI-Based Jamming Detection and Open-Set Classification for IIoT Networks},
      author={Aymen Bouferroum and Ildi Alla and Valeria Loscri and Abderrahim Benslimane and Vincent Lenders},
      year={2026},
      eprint={2606.22939},
      archivePrefix={arXiv},
      primaryClass={cs.CR},
      url={https://arxiv.org/abs/2606.22939},
}

This paper is the roadmap of my doctoral thesis: a lightweight, multi-layer, ML-based security framework for the Industrial IoT. It builds on the Tm-IIoT trust model and the Hybrid IIoT architecture, adds the Trust Convergence Acceleration approach for up to 28.6% faster trust decisions on degraded networks, and extends toward physical-layer threat detection and resilience against adversarial ML, aiming at real deployments on affordable, open-source hardware rather than simulation alone.

@article{bouferroum2026toward,
  title={Toward a Multi-Layer ML-Based Security Framework for Industrial IoT},
  author={Bouferroum, Aymen and Loscri, Valeria and Benslimane, Abderrahim},
  journal={arXiv preprint arXiv:2603.24111},
  year={2026}
}

Trust scores are only useful once they converge, and on real networks with fluctuating quality that can take far too long. TCA (Trust Convergence Acceleration) puts a Random Forest in the loop: it predicts how long convergence will take and adapts the trust model's transition probabilities on the fly. Under realistic Wi-Fi 6 conditions, TCA cuts trust-convergence time by up to 28.6% in challenging conditions while improving accuracy against malicious nodes.

@inproceedings{bouferroum2025accelerating,
  title={Accelerating Trust Convergence in IIoT: A ML Approach for Dynamic Network Conditions},
  author={Bouferroum, Aymen and Loscri, Valeria and Benslimane, Abderrahim},
  booktitle={IEEE Global Communications Conference (GLOBECOM) 2025},
  year={2025}
}

Drones in 6G networks need ultra-reliable, low-latency links while moving fast through three dimensions. We tackle spectrum allocation for aerial users in user-centric cell-free massive MIMO: a deep reinforcement learning agent (DDPG) assigns frequency subbands while balancing interference, spectral efficiency, and fairness under strict bandwidth limits. The learned policy adapts to altitude-varying conditions better than two bio-inspired optimizers (BGWO and BMFO) and shows the architecture can meet 6G performance targets for aerial networks.

@INPROCEEDINGS{11099147,
  author={Cheggour, Selina and Bouferroum, Aymen and Krishnan, Rahul},
  booktitle={2025 International Conference on Computer, Information and Telecommunication Systems (CITS)},
  title={Reinforcement Learning-Driven UC-CFmMIMO for UAVs: A Subband Allocation Framework},
  year={2025},
  volume={},
  number={},
  pages={1-8},
  keywords={6G mobile communication;Wireless communication;Reinforcement learning;Quality of service;Massive MIMO;Interference;Dynamic scheduling;Resource management;Vehicle dynamics;Radio spectrum management;Sixth-generation (6G);User-Centric Cell-Free Massive MIMO (UC-CFmMIMO);Aerial Vehicular Networks;Resource Allocation;Reinforcement Learning (RL);Deep Deterministic Policy Gradient (DDPG)},
  doi={10.1109/CITS65975.2025.11099147}}

FITNESS is a national research initiative funded by the French Research Agency (ANR) under France 2030, asking what still stands between the IoT and large-scale adoption. This white paper presents the project's vision and first findings on three fronts: IoT architecture and interoperability, the place of artificial intelligence in IoT, and energy efficiency, validated against concrete industrial use cases on the road to a secure, efficient, and interoperable IoT ecosystem.

@misc{cassiau2025fitness,
  author = {Cassiau, Nicolas and Achir, Nadjib and Adjih, Cédric and Andrieux, Guillaume and Bechkit, Walid and others},
  title  = {Overcoming the Technical Hurdles of IoT Adoption: the FITNESS Project Vision and Insights},
  year   = {2025},
  doi    = {10.5281/zenodo.17119689},
  note   = {hal-05257163},
  url    = {https://hal.science/hal-05257163}
}