With the proliferation of intelligent visual sensors, wireless camera networks are becoming a key research area for numerous applications, such as surveillance, object recognition and real-time scene analysis. However, optimizing data processing and transmission remains a major challenge due to bandwidth constraints, energy consumption and the limited computational capacity of embedded sensors. In a spirit of frugality, this thesis proposes an original distributed architecture based on federated artificial intelligence and a distributed ledger mechanism built on latent representations, in order to optimize information exchange within the network while ensuring data confidentiality and traceability.
To preserve the privacy of individuals, the cameras used are of very low resolution (32×32 pixels), thereby reducing the risk of identification while ensuring the collection of relevant information for the targeted applications. These cameras already exist and consist of a low-resolution sensor, an analog PIR sensor, a small FPGA (Max10 – low cost and low power) and an ESP32-C6 for managing network communication via Bluetooth.
Research problem
The use of wireless cameras in a dynamic environment imposes contradictory requirements. On the one hand, it is necessary to capture and analyze data to ensure optimal accuracy. On the other hand, energy consumption and transmission latency must be minimized. The existing hardware architecture, based on a small FPGA (Max10) and an ESP32 for Bluetooth communication, imposes strong constraints in terms of embedded processing optimization and exchange protocol efficiency.
To address these constraints, we propose a unified and original approach: each network node integrates an autoencoder whose latent space serves a dual purpose. On the one hand, it acts as a compressed representation of visual information, significantly reducing the volume of data to be transmitted over the Bluetooth network. On the other hand, each timestamped latent vector constitutes an entry in the distributed ledger, replicated across all nodes, allowing the reconstruction of the observation history by querying any node in the network. Federated AI then intervenes to align the latent spaces across different nodes, ensuring the interoperability of representations without requiring centralization of raw data.
Research objectives
The main objective is to design a unified architecture where the autoencoder embedded on each node serves as the cornerstone: it simultaneously ensures information compression, confidentiality preservation (latent representations are not visually interpretable) and the feeding of the distributed ledger.
This objective is broken down into four axes: designing quantized autoencoders (8 bits or less) capable of running on a Max10 FPGA while optimizing the trade-off between latent size, reconstruction quality and hardware resources; building a distributed ledger where each block is a timestamped latent, cryptographically chained and replicated across all nodes, with a potentially hybrid consensus mechanism (cryptographic combined with semantic coherence); implementing federated learning to align latent spaces across nodes and ensure their interoperability without ever exchanging raw images; and optimizing the Bluetooth protocol for exchanging latent vectors and federated update parameters under bandwidth and energy constraints.
Methodological approach
The methodology relies on a hybrid approach combining theory, simulation and experimentation on existing hardware, organized in four phases. The first phase is devoted to the design and implementation on the Max10 FPGA of quantized autoencoders (8-bit, 4-bit, binary) for 32×32 images, exploring efficient architectures and evaluating the trade-off between latent size, reconstruction quality and hardware resources. The second phase focuses on the distributed ledger: cryptographic chaining of latent vectors, Bluetooth replication protocol, management of the ESP32 memory constraints, history pruning mechanisms and formalization of the hybrid crypto-semantic consensus. The third phase is dedicated to federated learning for inter-node autoencoder alignment, with the evaluation of aggregation strategies adapted to network constraints and the study of latent space drift due to heterogeneous conditions across sensors. The fourth phase consists of integration on a real prototype followed by a global evaluation covering energy consumption, latency, AI accuracy, reconstruction quality from latent vectors and distributed ledger integrity.
Expected impact and contributions
The main contribution lies in the proposal of a unified architecture where the autoencoder’s latent space serves as the native support for the distributed ledger, merging compression, privacy protection and data traceability into a single operation. This is particularly suited to ultra-constrained nodes where each operation must serve multiple objectives. The thesis will also contribute to pushing the limits of federated learning on low-cost FPGAs and microcontrollers, a still largely unexplored area since the literature focuses on more powerful nodes such as smartphones or edge servers, with contributions on model quantization and heterogeneous latent space alignment. The custom-designed distributed ledger mechanism, with its hybrid cryptographic and semantic consensus, will open avenues applicable to other IoT sensor networks facing the same limitations. Finally, confidentiality is ensured at three complementary levels – low sensor resolution, exclusive exchange of visually non-interpretable latent vectors, and no centralization of raw data thanks to federated learning – which constitutes a coherent multi-layer privacy-by-design approach, paving the way for varied applications ranging from surveillance to autonomous object recognition in a context of frugality.
The originality of this thesis lies in the deep integration between embedded autoencoder, distributed ledger and federated learning within a wireless camera network with ultra-constrained resources. By making the latent space the vehicle for compression, confidentiality and distributed traceability alike, this architecture proposes a coherent and frugal solution to the challenges posed by intelligent visual sensor networks. The use of very low-resolution cameras, combined with decentralized collaborative learning, will reconcile performance, privacy protection and energy sobriety, while opening new perspectives for embedded technologies.
Thesis supervisor: François BERRY (PR) – Institut Pascal, francois.berry@uca.fr
Co-supervisors: François DELOBEL (MCF)-LIMOS, francois.delobel@uca.fr, Christophe BLANC (MCF- Institut Pascal), christophe.blanc@uca.fr
Funding: This thesis is funded by the MIAI Embed-AI research chair, whose goal is to propose hardware frugality solutions for AI.
To apply, please send a detailed CV, at least one recommendation letter from your Master’s supervisor, as well as transcripts from the last 3 years including your Master’s grades.
These documents should be sent to all three thesis supervisors.
