Subject Training

The Internet of Things (IoT) relies on a set of wireless communication technologies that connect sensors and devices constrained in terms of energy and computational capacity. To assess the quality of service of communications, performance indicators such as reliability (Packet Delivery Ratio, PDR), latency, and the number of retransmissions are commonly analyzed.

Today, the rise of machine learning opens new perspectives for network performance prediction. Rather than relying solely on analytical models or simulations, it is now possible to train predictive models using experimental or simulated data [3]. However, a major challenge lies in finding the right trade-off between model accuracy and lightweight design for deployment in constrained environments.

The concept of Knowledge Distillation (KD) [2] is a model compression technique that consists in transferring knowledge from a complex model (the teacher) to a smaller, lighter one (the student). This approach is particularly well suited to IoT networks, where it is essential to reduce memory and computation costs while maintaining predictive capability [1].

This internship aims to study the use of Knowledge Distillation for network performance prediction in wireless and IoT networks. Two approaches will be compared: ⬔ Global reduced model: a teacher model trained on the entire network to predict performance metrics (reliability, latency, retransmissions) based on configuration parameters (backoff, contention window, etc.), then distilled into a reduced student model. ⬔ Per-link models: individual models trained for each communication link to provide more fine-grained predictions.

The goal is to compare these two strategies in terms of predictive accuracy and computational cost. The identified tasks are as follows:

⬔ Literature review on Knowledge Distillation applied to IoT and wireless networks, ⬔ Familiarization with simulation (e.g., ns-3) and/or experimental environments (e.g., FIT IoT-Lab), ⬔ Collection of network performance data (reliability, latency, retransmissions) under different scenarios and configurations, ⬔ Training of teacher models (global and per-link), ⬔ Application of KD techniques to obtain reduced student models, ⬔ Comparison of both approaches in terms of accuracy and complexity.

References [1] Gad Gad et al. « Communication-Efficient Federated Learning in Drone-Assisted IoT Networks : Path Planning and Enhanced Knowledge Distillation Techniques ». In : 2023 IEEE 34th Annual International Symposium on Personal, Indoor and Mobile Radio Com- munications (PIMRC). 2023, p. 1-7. [2] Jianping Gou et al. « Knowledge distillation : A survey ». In : International journal of computer vision 129.6 (2021), p. 1789-1819. [3] Samir Si-Mohammed et al. « Ns+ ndt : Smart integration of network simulation in network digital twin, application to iot networks ». In : Future Generation Computer Systems 157 (2024), p. 124-144.