Hardware Design, Safety, and Security Dimensions of Spiking Neural Networks
The rising demand for low-power and real-time artificial intelligence challenges the limits of traditional computing architectures. Neuromorphic computing—drawing inspiration from the structure and operation of the human brain—has emerged as a promising alternative paradigm. At the heart of this approach lie Spiking Neural Networks (SNNs), which process information through discrete events, enabling asynchronous, energy-efficient computation well suited to embedded devices, autonomous systems, and sensory applications.
This thesis investigates the major challenges involved in transitioning SNNs from theoretical models to practical and deployable hardware systems. Unlike conventional AI models, SNNs operate with temporal dynamics and follow a non-von-Neumann architecture, calling for fundamentally different approaches in system design, testing, and security.
The work is organized around three main contributions:
- An innovative functional testing methodology specifically designed for SNN-based hardware architectures.
- The study of emerging security risks, including the development of untargeted adversarial attacks and a hardware Trojan attack for SNNs.
- A modular FPGA-based deployment framework enabling the transformation of abstract neural models into physical circuits.
Overall, this thesis strengthens the practical foundations of neuromorphic computing by enhancing the reliability, security, and accessibility of SNN hardware accelerators, bringing brain-inspired AI closer to real-world deployment.
Phd defence : 11/27/2025
Jury members :
Stefano DI CARLO, Politecnico di Torino, Italie [Rapporteur]
Elena Ioana VATAJELU, Université Grenoble Alpes (UGA), Laboratoire TIMA, France [Rapporteur]
Amir ZJAJO, Innatera Nanosystems B.V., Pays-Bas
Sylvain SAÏGHI, Université de Bordeaux, Laboratoire IMS, France
Haralampos-G. STRATIGOPOULOS, Sorbonne Université, Laboratoire LIP6, France