I work on machine learning for resource-constrained systems: making data from IoT devices, sensors and industrial machines smaller, safer and more useful.
Sensors and edge devices produce far more data than their batteries and radios can afford to send. I design lossy and lossless compression schemes, from DWT lifting and adaptive sampling to SZ-based and deep-learning approaches, that cut transmission cost while keeping the data useful for machine learning further down the line.
My PhD work focused on physiological signals collected by wireless body sensor networks. I continue to work on artifact filtering, compressed-domain classification and, more recently, how far vision-language models can go in reading ECGs without task-specific training.
Constrained devices need security that fits within their limits. I work on federated and cross-layer intrusion detection, anomaly detection for smart-grid networks, and lightweight cryptography and crypto-compression for multimedia IoT.
Industrial machines produce multivariate time series that call for early and reliable fault detection. My work here covers time-series classification with CNNs and transformers, priority sampling, and semantic interoperability between heterogeneous industrial systems.
Across all of these themes I look for models that are small enough to deploy and transparent enough to trust. That includes knowledge distillation, frugal object detection, privacy-preserving federated learning and, more recently, LLM agents grounded in knowledge graphs.
FEMTO-ST Institute, CNRS, DISC Department, AIMOS team.
Computer Science Department, IUT Nord Franche-Comté, Université Marie et Louis Pasteur (UMLP).
