Lieu : Remote or Grenoble or Lyon · Contrat : Stage · Rémunération : 1000€/m brut €
Ryax Technologies est une startup dont l'objectif est d'automatiser les analyses de données dans les entreprises. Nos clients vont de la pharma au digital marketing en passant par l'industrie lourde.
Pour cela nous avons développé une technologie permettant de créer des chaînes de traitements rapidement et intuitivement et d'ensuite exécuter ces chaînes sur n'importe quelle infrastructure informatique.
Adoption of Machine Learning (ML) and more specifically Deep Leaning (DL) across different industries has helped automate tasks in novel ways for businesses. However, shifts in the way we use data to solve problems necessitates a deeper understanding of the shortcomings in our systems. Previously human tasks now fall within the scope of our algorithms, which have been shown to behave in unexpected and unfair ways. Not only can these systems be exploited by adversaries, but systems reliant on data may also have more fundamental drawbacks such as data bias and lack of privacy.
Operating DL systems in real world scenarios such as surveillance, autonomous vehicles, and online recommendation systems puts the private information of human beings at risk, and in certain scenarios can result in direct personal harm should the algorithm fail. This topic focuses on notions of security and privacy in ML systems. Adversarial Robustness (AR) will be studied from the attack and defense standpoints to understand the inner workings of Deep Neural Networks (DNNs), and how they can be both exploited and protected. Further, the internship will address the notion of differential privacy in scenarios that use private data, and how this effects ML integrity.
This internship is hosted by Ryax Technologies; a startup in Lyon, France providing data engineering tools for businesses to develop and scale their data science and ML operations on distributed production environments. The goal for the student is to gain meaningful understanding of how the topics above pervade in settings relevant to the use-cases at Ryax, and to gain hands-on engineering experience relevant to this line of work. To this end, the student has the opportunity to create workflows within the Ryax platform to address this field of research. Some examples of relevant workflows are;
- Adding noise to obfuscate data in privacy-intensive scenarios, and de-noising to assess risk.
- Threat assessments on ML models with state of the art adversarial attacks.
- Generating differentially private datasets from user data + distribution analyses of informative priors.
The internship will begin with a guided review of the state of the art in Robust Deep Learning, differential privacy, and the interplay between these topics. Then, the student has liberty to decide which aspects they would like to study more rigorously. Finally, the intern will conduct meaningful experiments and culminate their work with a masters thesis.
- Basic ML concepts e.g gradient descent, regression, representation learning, and some basics of deep learning
- Intuition on Convex Optimization, such as sets, matrix properties
- Python programming language (knowledge in other languages such as C, bash, a plus)
Technologies to be used (no prior knowledge needed, but is a plus):
- Docker & Kubernetes
- Adversarial Machine Learning
- Sparse Matrix Methods
Contact: Yiannis Georgiou & Charles Marshall Mail: firstname.lastname@example.org - email@example.com