Privacy-preserving federated learning over vertically partitioned data from heterogeneous participants

Federated learning enables multiple participants to collaboratively train a global model, without sharing their data, but only model parameters are exchanged between the participants and the server. In vertical federated learning (VFL), datasets of the participants share similar samples, but have different features. For instance, companies and institutions from different fields own data with different features of overlapping samples collaborate to solve a machine learning task. Though data are private, VFL remains vulnerable to attacks such as label and feature inference attacks. Various privacy measures (e.g., differential privacy, homomorphic encryption) have been investigated to prevent privacy leakage. Choosing the appropriate measures is a challenging task as it depends on the VLF architecture and the desired level of privacy (e.g., local models, intermediate results, learned models). The variability of each participant’s system can also result in high latency and asynchronous updates, affecting training efficiency and model effectiveness.

The aim of this thesis is to propose methods to enable privacy-preserving VFL, taking into account the heterogeneity of the participants. First, the candidate will study the architectures of VFL models and the privacy measures to propose privacy-preserving protocols for VFL. Second, the candidate will investigate the impacts of the heterogeneity of the participants’ system such as computation and communication resources to devise solutions to render the VFL protocols robust to such heterogeneity. Third, the trade-offs among effectiveness, privacy, and efficiency in VFL will be explored to propose a practical framework for adjusting the protocols according to the requirements of a given machine learning problem.

Environmental applications of the metrological study of photonuclear reactions on light elements.

Through a thesis and a project funded by the LNE, the LNHB is developing a prototype for the detection of illicit materials using the active photon interrogation method, based on the spectrometry of photoneutrons emitted by targets irradiated by a linear electron accelerator. This new thesis topic involves studying photonuclear reactions on light elements for environmental applications, primarily through a better understanding of photoneutrons, secondary neutrons and cosmogenic radionuclide production rates in the atmosphere during terrestrial gamma ray flashes associated with thunderstorms or gamma-ray bursts of cosmic origin. The LNHB-MD's unique experimental facility will be use to obtain basic nuclear data, such as the angular and energy distributions of photoneutrons emitted by light elements, and to characterize activation products. The data collected will be use to improve the description of photonuclear processes for light elements in Monte-Carlo codes, and to estimate their influence on measurable quantities through simulation codes for environmental phenomena. The same methodology can be apply to the study of photonuclear reactions in rocks - composed mainly of light elements - following irradiation by high-energy photons of natural or artificial origin.

Simulation of ultrasonic waves interaction phenomena with metal microstructure for imaging and characterisation purposes

The interaction of waves with matter is highly dependent on the frequency of these waves and the scale of their wavelengths in relation to the properties of the medium under consideration. In the context of ultrasound imaging applications, particularly concerning metals, the scales considered are generally on the order of a millimeter (ranging from a tenth to several tens of millimeters). However, depending on the manufacturing processes used, metallic materials, often anisotropic, can also have a microstructure with heterogeneities of similar dimensions. Consequently, ultrasonic waves propagating through metals may, under certain circumstances, be significantly influenced by their microstructures. This can either pose limitations to certain ultrasonic techniques (e.g., attenuation, structural noise) or provide an opportunity to estimate local properties of the inspected metal.
The general objective of the proposed thesis is to gain a deeper understanding of the link between microstructure and ultrasonic wave behaviour for large classes of material by benefiting from the combined knowledge of LEM3 for the generation of virtual microstructure and of the CEA for the simulation of ultrasonic wave propagation.
This work will combine the acquisition and analysis of experimental data (material and ultrasound), the use of simulation tools, and the statistical processing of data. This will enable an analysis of behaviors based on material classes, and possibly the implementation of inversion procedures to characterize a microstructure from a set of ultrasonic data. The combination of these methods will enable a holistic approach, contributing to significant advances in this field.

architecture for embedded system of Automated and Reliable Mapping of indoor installations

The research focuses on the 3D localization of data from measurements inside buildings, where satellite location systems, such as GPS, are not operational. Different solutions exist in the literature, they rely in particular on the use of SLAM (Simultaneous Localization And Mapping) algorithms, but the 3D reconstruction is generally carried out a posteriori. In order to be able to propose this type of approach for embedded systems, a first thesis was carried out and led to a choice of algorithms to embed and a draft of the electronic architecture. A first proof of concept was also realized. Continuing this work, the thesis will have to propose a method allowing the localization device to be easily embedded on a wide range of nuclear measuring equipment (diameter, contamination meter, portable spectrometry, etc.). The work is not limited to a simple integration phase; it requires an architectural exploration, which will be based on adequacy between algorithm and architecture. These approaches will make it possible to respect different criteria, such as weight and small size so as not to compromise ergonomics for the operators carrying out the maps and quality of the reconstruction to ensure the reliability of the input data for the Digital Twin models.

Development of x-ray phase contrast and dark field imaging numerical model

Since 2013, CEA List (Université Paris Saclay) has been developing phase contrast X-ray imaging methods, in particular using multi-lateral shearing interferometry. In addition to absorption information, the phase shift of X-rays provides additional contrast and sensitivity on the image, particularly for materials with low atomic numbers or low density.
Various techniques have been developed to generate a phase contrast, based in particular on the addition of a random or regular intensity modulator (sandpaper or grid). In addition, dark field imaging has emerged as a valuable complementary signal to phase contrast imaging. The dark field signal comes from the small-angle scattering of fine structures in the sample. In particular, the dark field signal has proven it sensibility to reveal features of the sample that remain invisible by conventional means. It can, for example, reveal the microstructural properties of the lung in cases of chronic obstructive pulmonary diseases.
The continuation of these developments requires the implementation of a numerical model producing sufficiently accurate images that are representative of an experimental system.
The aim of the thesis is to develop a numerical model that takes into account the phenomena of phase contrast and scattering, in particularby refraining from a classic modelling hypothesis, which is the consideration of an thin object (projected thickness hypotheseis). Failure to take this assumption into account will have to be dealt with in order to move towards phase imaging on a thick object (e.g. a thorax).
As a general rule, phase contrast is represented using models based on wave propagation. In contrast, scattering phenomena are usually simulated using a particle-based approach, often using Monte Carlo techniques. In this study, a combined approach will be developed with experimental validation.
The thesis will be carried out in CEA List with people who have solid numerical and experimental skills.

Topology reconstruction of a ramified network by multisensor reflectometry

Smart Grids aim at monitoring and controlling electric power networks. Many parameters have to be monitored such as production and consumption units, and the integrity of the structure of the interconnection netwotk itself.

Smart grids aim at enhancing the quality of service while protecting people and infrastructures. In this area of research, most algorithms are deployed for taking the human out of the retroaction loops in order to maximize the availability and the reactivity. For that reason, artificial intelligence based algorithms are increasingly incorporated in decision loops.

In that industrial context, we are interested in methods that aim at estimating electrical network topologies. The topology of a network includes the length of the cable lines and their electrical properties, so as the characterictics of the loads that are connected to the networks (production and consumption units), and also potential faults in the network. In the end, the accurate estimation of the topology may be used to monitor the network with more accuracy with the help of a more accurate a priori information.

In order to characterize the topology, we propose to deploy either a single or a distributed set of electric reflectometers. These devices inject signals in the network under test and the study of the reflections gives information back which can be used to reveal the structure of the network. More precisely, every impedance discontinuity along the line wil cause partial reflections of the waves.

Previous works were conducted by our team of researchers on that topology reconstruction topic, by exploiting optimization algorithms coupled to a simulator. We would like to extend these works in two directions. First, we would like to explore a machine learning regressor-based approach in a mono sensor version. Second, we would like to estimate the topology by combining the measurements from multiple sensors, either with already available optimization-based approachs, or by the new machine learning-based approach.

Development of a neutron/gamma coincidence measurement system for the characterization of radionuclide neutron sources

This PhD work is part of sources calibration activities at the LNHB-MA and R&D activities within the SIMRI aimed at developing neutron measurement systems for the CEA and the nuclear industry. The objective of the PhD work is to develop a new measurement system using neutron/gamma coincidences to enable the characterization of the (alpha,n)-type neutron sources. These sources consists of a homogeneous mixture of an alpha particle emitter and the target substance, the nuclei of which emit neutrons via a nuclear reaction. As for example, we can cite for example: AmBe, PuBe, CmBe, or even exotic source of high emissivity and mixing several alpha radionuclides (ex. AmPuBe). For this familly of sources, the emission of neutron by reaction (alpha,n) is in simultaneous cascade with a characteristic gamma at 4.4 MeV. The detection of the neutron and the gamma in coincidence is likely to provide information of interest in the source characterization in terms of emission rate and spectral fluence. The objective is to measure precisely gamma and neutron signatures as well as gamma/neutron intensity ratios resulting from the nuclear reaction. The new measurement device must also be able to measure neutrons emitted by the spontaneous fission reaction or by (n,2n) reaction in beryllium. Others photon emission can be also provide information of interest, ex. the emission of a gamma at 2.2 MeV resulting from the capture on hydrogen. The neutron/gamma coincidence measurements can be also used to improve the evaluation of nuclear data such as cross sections of certain elements, ex. (n,gamma) reaction on oxygen or hydrogen.

Inscription of Optical Waveguide in Silica or Saphirre Optical Fbers and Characterization at High Temperature

Fiber Bragg Gratings are structures, photo-inscribed with femtosecond laser, inside the fiber core and can be used as band pass optical filters (centered around the Bragg wavelength). Bragg wavelengths are easily multiplexed and they give to us the necessary information. Silica-based fiber Bragg gratings are point sensors and can measure temperature up to 1200°C. For higher temperature, sapphire based optical fiber are used, since they can withstand temperature up to 2000°C. However, the sapphire optical fiber is coreless, which leads to an extreme multimodal behavior. Consequently, the measure is less precise and the signal-to-noise ratio is low, compare to a classical silica-based grating. Moreover, each modification of the fiber surface, change the grating spectra.
The thesis objective is the creation of an optical waveguide inside the sapphire fiber, which will leads to less propagation modes inside the fiber, in order to obtain new perspectives for the monitoring in high temperature environments (airplane engines, nuclear reactors, …), which is one of the missions of the DRT/LIST-DIN. To obtain this result, the photo-inscription of a cladding is necessary: the cladding will be ring shape and the internal diameter – i.e. the core – will be few tenths of micrometers. Other techniques are also investigated, such as ion implantation, to create an amorphous sapphire cladding. Then these new structures will be characterized up to 2000°C and under high dynamic pressure (> 10 GPa).

Differentiable surrogate for simulation-based inference

Many models of complex phenomena (physics, molecular dynamics, etc.) have no
global analytical expressions but admit implementations in silico in form of
forward simulators. In turn, forward simulations are used to solve inverse
problems: given observations of the phenomena find its initial conditions
viewed as input parameters of the simulator.

In statistical terms solving such an inverse problem corresponds to sampling
from a (bayesian) posterior distribution with the implicit likelihood given by the
simulator - this provides (at least) an answer to the problem with
error bounds in form of uncertainty estimates. High dimensionality and/or
computational load hinder use of simple classical methods (like ABC or
Kernel Density Estimator) and lead to construction of surrogates that approximate
the intractable likelihood coupled with amenable schemes for posterior sampling.

Recent advances in Automatic Differentiation models allow construction of
such surrogates which are yet in their early development. In this thesis
we aim to study and develop new ways to construct differentiable surrogates
and apply them on a number of realistic problems starting
with a number of applications in nuclear imaging.

Study and exploitation of Barkhausen noise spectral information for the characterization of steels

The use of magnetic Barkhausen noise (MBN) measurements for assessing the structural health of magnetic materials has become an important industrial technique the last years. The interest in the application of this technique stems from the strong dependence of the MBN signals on the material microstructure as well as its stress level and its chemical composition.
The development of robust and reliable analysis tools based on MBN signals is however greatly impeded by the complexity of the underline physics and its sensitivity upon the details of the microstructure. Although a number of models has been proposed in the last decades and significant progress has been reported in terms of the understanding of the phenomenon, a complete theory is still lacking.
Due to this lack of understanding and the complexity of the MBN signals, the current state of the art from the non-destructive testing (NDT) perspective is almost entirely based on the measurement and the analysis of the signal envelope. The spectral information although rich in content is ignored at this level. Yet, it has been demonstrated that the MBN spectrum can give rise to classification of the magnetic materials at different universality classes based on microstructural features, notably the degree of disorder.
The proposed Ph.D. aims to contribute in the use of spectrum measurements for the characterisation of magnetic materials, notably steels. Accurate MBN measurements obtained from different microstructures using a dedicated setup (developed in the context of the Ph.D. work) will be analysed and compared with theoretical simulations based on tools previously developed by the host institute in order to
• Validate and fine-tune the theoretical models
• Study the impact of the microstructure (grain size, dislocations) to the spectrum features
• Explore the classification of the considered microstructures in different classes
Starting from well-known model materials (FeSi and FeCo), for which a great amount of published results exist and hence can be used as reference, the study will be then focused on some important industrial steel grades like the interstitial-free (IF) and low-carbon (LC) steels.
The proposed Ph.D. thesis will be jointly directed and supervised by the French atomic and alternative energies commission (commissariat à l'énergie atomique et aux énergies alternatives, CEA) and the CEIT Institute. The main part of the work will be hosted at the CEA research centre at Saclay, France with possible stays at the CEIT institute in San Sebastian, Spain.
The sought candidate profile is compatible with physicists and engineers with a good background in solid-state physics and a solid understanding of electromagnetism. Basic metallurgical notions and a familiarisation with standard laboratory equipment is also expected. Basic programming knowledge will be helpful. The candidate is also assumed to have good communication skills in English.
The candidate will benefit from access to the experimental facilities of both centres, the central CEA library and the CEA transport network as well as the restaurant facilities.