© Robin Gerzaguet.
© Robin Gerzaguet.
You can find here some aspects of my current or past research activities mostly on digital communication systems and architectures and software defined radios. Do not hesitate to reach me to have further informations on these projects.
I am always looking for highly skilled and passionate graduate students in those domains so if you feel that your are the special one, do not hesitate to contact me.
My work falls within the framework of deploying a sensor network that collects, processes, and transmits environmental or contextual data from sensors deployed over a given area.
Let’s formulate the problem simply: to perform a processing task, an algorithm must be designed and then integrated into a system full of constraints. This requires an alignment between the algorithm and the target architecture. This is the core of my research topic : we call that algorithm-architecture adéquation (and we think it's beautiful)
This theme requires different areas of work where my contributions lie:
1️⃣ Design of signal processing algorithms, for telecom applications in the broad sense or for digital signal processing acquired by sensors.
2️⃣ Rapid prototyping methodology, applied to software-defined radio (SDR).
3️⃣ Implementability of signal processing solutions on embedded targets.
These research axes are supported by various collaborative projects and span different application contexts. My three main contexts include:
RF security 🔒.
Rapid prototyping using the Julia language 💻.
Frugal (or lightweight) IA
For a more precise and Triskell-based view of my research themes, see the figure below.
In the field of communication and information exchange, digital technologies dominate, and modern systems use digital signal formats. Adapting the communication protocol means working on all layers of the OSI model. My work has focused on the first layer of the OSI model, the physical layer (sometimes called L1 for Layer 1) in the context of transmissions over fiber and wireless systems.
This co-design must be aligned with the node's constraints (energy, latency, throughput, processing capacity), but also with the transmission constraints. Specifically, the propagation channel (fiber or air) and the hardware imperfections of the components involved in transmission and reception can degrade the quality of the decoded signal and introduce errors.
Several strategies can be implemented in the digital transmission chain to mitigate these phenomena:
Signal shaping: Make it more robust to disturbances — this is known as waveform design.
Compensation algorithms: Implement digital algorithms to compensate for these imperfections and disturbances.
Leverage imperfections: Use the uniqueness of each node’s hardware imperfections for identification purposes.
A key aspect of my work is the confrontation with hardware implementations and real signals. After joining IRISA, this inclination evolved into a dedicated research axis, alongside the deployment of the EEWOK Rose software-defined radio platform.
Thus, if axis 1 focuses on methods and algorithms with formal signal processing, axis 2 is about algorithm-architecture adequacy for communication systems based on SDR. SDRs are radio transmission architectures in which the analog part is minimized, with most of the processing done via software, offering great flexibility.
This area of research revolves around two aspects:
Design of processing blocks and SDR architectures: The algorithm-architecture match requires optimizing algorithms for a target architecture.
SDR programming methodology: Working on the expressiveness-performance tradeoff by using high-level programming languages, particularly Julia.
The last axis is both a guiding thread in implementing my work on hardware targets and an emerging area (since 2024). It coincides with the start of several collaborative activities around embedded processing.
An advantage of deploying a sensor network with local processing is the possibility of integrating intelligence, particularly machine learning. However, integrating these techniques introduces many challenges due to the constraints of embedded systems: memory management and energy management.
My work in this emerging field focuses on two aspects:
Implementation of lightweight neural networks: Combining quantization and pruning approaches to design lightweight networks and taking into account energy as a key indicator for performance.
Design of electronic boards and algorithm-architecture matching: On these embedded platforms, often based on microcontrollers, with limited memory and 16-bit instruction sets, suitable execution structures need to be developed.
