In today's connected world, we heavily rely on network protocols to communicate with one another. To ensure the confidentiality and integrity of the exchanged data, we need to assess and improve the security of these protocols and their implementation. This observation obviously concerns so-called security protocols, such as TLS and SSH, but it is also important to consider high-level, complex, application-level protocols, such as HTTP/2 and lower-level protocols such as DNS or BGP. Security flaws are indeed pervasive in network protocols, at the specification level or due to implementation flaws: incomplete specification, memory corruption bugs, logical errors, state machine shortcuts, cryptographic attacks, etc. The GASP project proposes a generic framework to describe protocols and automatically derive tools. We will first define a description language for messages, and derive parsers and scanning tools. Then, we will design languages to describe state machines, from which reference implementations can be derived. Finally, we will develop protocol fuzzers and test existing implementations. Our goal is thus to provide a generic approach to secure protocols, from observation to implementation to testing. The short-term impact of the project would be to improve the knowledge about protocol deployment and implementations. We will also aim at sharing the produced tools as well as the data collected and results obtained during the project. In the long run, the proposed languages could help produce better specifications and improve the time of development for protocol stacks.

Runtime systems have to take decision that are critical for the performances of parallel applications. Unfortunately, these decisions can only use heuristics based on the current status of the application in order to estimate how it will behave in the future. As a consequence, runtime systems may take decisions that degrade performances instead of improving them. pythia aims at providing runtime systems with means to accurately predict the future. For this, pythia relies on the deterministic nature of most parallel applications: most programs will behave similarly from one run to another. Thus, we will design a tool-chain that analyzes the execution of a program to provide hints to the runtime systems during future executions of the same program. Thanks to these hints, a runtime system could base its decisions on both the current status of the application, and the future behavior of the program.

We are interested in reflected stochastic processes involved in several systems of queueing networks. These stochastic models have been developed due to their numerous applications in operations research, risk theory, telecommunication, data sciences and also in population biology. The issues raised by such stochastic systems are intertwined with many fields of mathematics (probability, complex analysis, combinatorics, differential Galois theory). A typical example of such a continuous process is obliquely reflected Brownian motion in an orthant. Their discrete analogues, random walks in cones, are also very famous processes. There are a wide variety of interesting questions regarding these reflected processes which can be either recurrent or transient. Among these questions there are the study of invariant measures, Green's functions, time to reach an edge or the vertex, algebraic nature of the generating function, harmonic functions and Martin boundary associated to the process. A now standard method to study these problems is to establish kernel functional equations involving generating functions. The project intends to decompartmentalize certain approaches by mixing three different techniques used to solve these functional equations: Boundary value problems, Tutte's invariant approach and q-difference equations. The project's objectives are to study: 1. Persistence, extinction and quasi-stationary distribution in a cone. 2. Ruin and escape for transient process in an orthant. 3. Monte Carlo simulation for Brownian motion in cones. 4. Additive and multiplicative decoupling for Tutte's invariant approach. 5. Transition kernel and space-time functional equation. The funding would be mostly used to fund two years of postdoctoral fellowship(s) and to organize an international conference. The scientific coordinator's team will be made up of four complementary dynamic young researchers: Pierre Bras, Thomas Dreyfus, Andrew Elvey Price and Sofia Tarricone.

Transfer Learning (TL) aims to enhance the learning process of a new target task by leveraging knowledge gained from previous executions of different yet similar source tasks. Offering superior performance compared to traditional machine learning methods, especially in scenarios of data scarcity, diversification search, or frugality, TL is gaining increasing traction within both the statistical learning community and among industrial practitioners. A wide variety of statistical and algorithmic methods have been developed to address the questions of "what" (parameters, characteristics, etc.) to transfer from the source to the target and "how" to do it. Typically, the efficiency of these methods is assessed through numerical experiments, with only a handful accompanied by theoretical guarantees. Approaching the TL problem from a different perspective, DECATTLON aims to develop mathematical tools to address the following three issues: (1) Quantitatively measure transferability to determine the relevance of employing transfer in different frameworks (non-parametric regression, classification, domain adaptation), (2) Select the most advantageous sources in multi-source domain adaptation and define the optimal oracle calling strategy in a problem combining TL and active learning, (3) Develop TL methods in models evolving over time: PINNs, stochastic differential equations, Markov decision processes with application to solving a bin packing problem. DECATTLON's tool performance will be rigorously evaluated quantitatively, with accompanying theoretical guarantees. Moreover, these tools will undergo real-world testing and their applicability for potential industrial transfer will be thoroughly discussed.

The health sector is currently challenged by the advent of e-health services and by exceptional circumstances, e.g., the Covid-19 pandemic. These challenges have mainly showed a critical need on the wide collaboration between health actors and the efficient sharing of concrete data. However, the processing of medical information, considered as highly sensitive data, requires the compliance with a strict national and international legal framework. Encompassing these challenges requires future health infrastructures and services that go beyond their traditional rigidity by integrating privacy-preserving collaborative data-sharing environments, to reach equitable societal and health gains. EQUIHid will accelerate the digitization of data-centric health services, through the re-design of equitable Artificial Intelligence (AI) collaborative approaches for a privacy-preserving processing of medical information.