Topics
The postdoc focuses on novel and advanced embedded AI, combining two complementary aspects. On the one hand, the TinyML aspect [1], focused on the implementation of AI directly on constrained microcontrollers. On the other hand, privacy-preserving Machine Learning techniques using split computing [2,3], which involves offloading partial execution (inference) of a neural network from a client to server.
Depending on profile of the applicant, the topic will be tilt towards one aspect or the other (ideally both would be combined).
TinyML aspects: The goal is to implement efficient AI model execution (TinyML) on microcontrollers, and manage AI models (MLOps: remote updates, performance monitoring – here secure TinyMLOps) on hardware such as Nordic nRF52, STM32, ESP32, or RISC-V (while networking technologies include BLE, 802.15.4, or LTE-M). On top of this hardware, prototypes will be developed in conjunction with an open-source operating system written in embedded Rust (Ariel OS [4]) or embedded C (RIOT OS [5]). These prototypes will be co-developed and tested with Freie Universität Berlin. This project follows up on Ariel-ML [7] and RIOT-ML [6], also applied to concrete
use cases.
Privacy-preserving ML aspects: Split computing involves distributing the execution (inference) of a neural network between a client and a server by splitting the model at an intermediate layer. This approach is particularly useful for embedded systems, edge computing, and resource-constrained devices, as it reduces local computational costs, energy consumption, or latency. However, it also raises privacy concerns: the intermediate representations transmitted to the server may contain enough information to allow partial reconstruction of the input or inference of sensitive attributes. Recent state-of-the-art reviews such as [3] present the main mechanisms of split computing, their system benefits, and the associated open challenges. Other recent work such as [8] shows that intermediate representations can indeed be exploited in reconstruction attacks and proposes experimental frameworks to evaluate these leaks. Other work focuses on defense mechanisms aimed at reducing the sensitive information contained in these representations, while preserving performance on the target task [2]. We will study the performance of split computing with a privacy-preserving angle, and implement experimental prototypes evaluated on heterogeneous embedded system hardware.
[1] Capogrosso, L., Cunico, F., Cheng, D.S., Fummi, F. and Cristani, M., 2024. "A machine learning-oriented survey on tiny machine learning". IEEE Access, 12, pp.23406-23426.
[2] Ruijun Deng, Zhihui Lu, and Qiang Duan. “InfoDecom: Decomposing Information for Defending Against Privacy Leakage in Split Inference”. In: Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40. 25. 2026, pp. 20737–20745.
[3] Yoshitomo Matsubara, Marco Levorato, and Francesco Restuccia. “Split Computing and Early Exiting for Deep Learning Applications: Survey and Research Challenges”. In: ACM Computing Surveys 55.5 (2022), pp. 1–30
[4] Ariel: https://ariel-os.org
[5] RIOT: https://riot-os.org
[6] Huang, Z., Zandberg, K., Schleiser, K., & Baccelli, E. (2025). RIOT-ML: toolkit for over-the-air secure updates and performance evaluation of TinyML models. Annals of Telecommunications, 80(3), 283-297.
[7] Z. Huang, G. Myung, K. Schleiser, E. Baccelli. Ariel-ML: Computing Parallelization with Embedded Rust for
Neural Networks on Heterogeneous Multi-core Microcontrollers. Preprint: https://arxiv.org/pdf/2512.09800v1
[8] Abhishek Singh et al. “SIMBA: Split Inference - Mechanisms, Benchmarks and Attacks”. In: European Conference on Computer Vision (ECCV). 2024.
Responsibilities
The researcher will be responsible for the design and development of the conceptual parts (AI, model, protocols), the use of datasets, implementations with application to a concrete use involving experiments on embedded hardware. he recruited researcher will interact with Inria scientists (including members of TRiBE and/or
PETSCRAFT teams) in the fields of machine learning and secure low-power IoT communication protocols, as well as the open-source developer communities of Ariel OS / RIOT, including our partners at Freie Universität Berlin, and engineers we collaborate with at Campus Cyber through partnerships with Orange and La Poste on real-world use cases (depending on the use case).
Coordination/Management
The recruited person will be the main point of contact between Inria, Freie Universität Berlin, the maintainers of Ariel OS and/or RIOT, including software engineers we collaborate with at Campus Cyber, and last but not least, the involved industrial partners deploying the use case.
