Activez les alertes d’offres d’emploi par e-mail !

PhD Position F / M Privacy-Enabled AI Job Execution on Heterogeneous Consumer Hardware Architectures

INRIA

Nantes

Sur place

EUR 80 000 - 100 000

Plein temps

Il y a 11 jours

Générez un CV personnalisé en quelques minutes

Décrochez un entretien et gagnez plus. En savoir plus

Repartez de zéro ou importez un CV existant

Résumé du poste

Une opportunité de doctorat au sein de l'INRIA à Nantes, axée sur le développement d'un middleware pour exécuter des charges de travail d'IA sur des ressources informatiques décentralisées. Le candidat idéal possédera des compétences solides en cybersécurité et en programmation, avec un intérêt pour la recherche appliquée et la collaboration entre le milieu académique et l'industrie. Le poste propose une rémunération compétitive, divers avantages sociaux et la possibilité de télétravail après six mois.

Prestations

Remboursement partiel des frais de transport public
7 semaines de congés annuels + 10 jours de RTT
Possibilité de télétravail
Équipement professionnel disponible
Événements sociaux et culturels
Accès à la formation professionnelle
Couverture de sécurité sociale

Qualifications

  • Une bonne compréhension des modèles de menaces et de la gestion des clés.
  • Une expérience en programmation système et en AI est un plus.

Responsabilités

  • Conduire des recherches originales sur la sécurité dans l'AI.
  • Collaborer avec l'équipe STACK et les partenaires de Hive.
  • Développer un middleware pour distribuer les charges de travail AI.

Connaissances

Compréhension des principes de cybersécurité
Programmation système (C / C++, Rust)
Langage de haut niveau (Python)
Cadres AI (PyTorch, Tensorflow)

Description du poste

The position is part of a collaboration between the Hive company and the STACK team. The successful candidate will be part of the STACK team based at IMT Atlantique, Nantes.

About STACK : STACK is a research group focusing on challenges related to the management and advanced usages of Utility Computing infrastructures (i.e. Cloud, Fog, Edge, and beyond). More specifically, the team is interested in delivering appropriate system abstractions to operate massively geo-distributed ICT infrastructures, from the lowest (system) levels to the highest (application development) ones, and addressing crosscutting dimensions such as energy or security. Those infrastructures are critical for the emergence of new kind of applications related to the digitalization of the industry and the public sector (aka Industrial and Tactile Internet).

About the Hive company : Hive is shaping the future of cloud computing by leveraging unused computing capacity to provide a decentralized, environmentally friendly, and user-empowered alternative to traditional cloud services.

Context

Taking security into account in distributed AI work is essential to respect user privacy, be it for inference or model training [3]. This is particularly the case when using consumer computers for computing tasks, a technique known as Desktop Grid Computing [1]. This otherwise idle hardware enables low-cost execution of computing tasks. However, in the AI case, it brings privacy issues as personal training or inference data is co-located with other users’ processes. One way to overcome this challenge is to employ the hardware security capabilities provided by modern processors, coprocessors, and chipsets [6,2]. The security features present in modern consumer computers include, for example, the Trusted Platform Module (TPM) chipsets integrated in most x86 motherboards, or Microsoft Pluton [5] in modern x86 processors. These modules provide numerous functions to help secure applications, such as key storage, encryption, and decryption. In addition recent processors have capabilities to securely run code in isolation, using for instance Intel SGX, SEV or TrustZone technologies [7]. TrustZone has been used successfully on AI projects in recent work [2, 4], as ARM platforms are becoming increasingly widespread in consumer computers and data centers. However, not all these security capabilities are identical, nor do they provide the same security guarantees depending on the family and model of processors and chipsets employed. Added to this is the complexity of using this type of architecture, which often requires specific programming models resulting in heavy adaptations of the executor code [4], or in other cases a very reduced performance [3].

Goals

The main objective of this PhD is to provide middleware-like software support able to distribute and run AI workloads on resources hosted on end-user computers, according to their software and hardware capabilities, while taking into account users’ privacy requirements. Several research questions will be considered :

  • How to ensure the privacy of AI processing on computers with heterogeneous capabilities, taking advantage of end-user hardware security capabilities?
  • How to adapt AI processing to take advantage of these security capabilities?
  • What trade-offs between security and performance need to be addressed?
  • How to dynamically distribute jobs across participating nodes to balance those trade-offs?

The development of this middleware is based on the following steps : First, an analysis of the capabilities and limitations of TPM and TrustZone will be carried out to understand how these technologies can secure AI processing. Then, the middleware will be designed to automatically detect the hardware capabilities of computers and adjust the security level according to the available configuration. To achieve this goal, it will integrate mechanisms to isolate AI processing with TrustZone and secure the storage of sensitive data with TPM. Finally, we will develop an AI job orchestrator based on optimization models and / or heuristics to determine the most efficient deployment according to security, performance, and machine availability needs. Experiments on devices equipped with these technologies will validate the middleware. The tests will measure the impact of dynamic adaptation on security and performance, providing data on the feasibility and effectiveness of the middleware in real-world scenarios.

1] Evgeny Ivashko, Ilya Chernov, and Natalia Nikitina. “A Survey of Desktop Grid Scheduling”. In : IEEE Transactions on Parallel and Distributed Systems 29.12 (Dec. 2018), pp.2882–2895. issn : 1558-2183. doi : 10.1109 / TPDS.2018.2850004. url : (visited on 04 / 29 / 2025).

2] Aghiles Ait Messaoud et al. “Shielding federated learning systems against inference attacks with ARM TrustZone”. In : Proceedings of the 23rd ACM / IFIP International Middleware Conference. Middleware ’22. New York, NY, USA : Association for Computing Machinery, Nov. 2022, pp. 335–348. isbn : 978-1-4503-9340-9. doi : 10.1145 / 3528535.3565255. url : (visited on 11 / 05 / 2024).

3] Fan Mo, Zahra Tarkhani, and Hamed Haddadi. “Machine Learning with Confidential Computing : A Systematization of Knowledge”. In : ACM Comput. Surv. 56.11 (June 2024), 281 : 1–281 : 40. issn : 0360-0300. doi : 10.1145 / 3670007. url : (visited on 10 / 14 / 2024).

4] Arttu Paju et al. “SoK : A Systematic Review of TEE Usage for Developing Trusted Applications”. In : Proceedings of the 18th International Conference on Availability, Reliability and Security. ARES ’23. New York, NY, USA : Association for Computing Machinery, Aug. 2023, pp. 1–15. isbn : 9798400707728. doi : 10.1145 / 3600160.3600169. url : (visited on 10 / 17 / 2024).

5] Vinay Pamnani. Microsoft Pluton security processor. en-us. July 2024. url : (visited on 11 / 05 / 2024).

6] Lianying Zhao et al. “A Survey of Hardware Improvements to Secure Program Execution”. en. In : ACM Computing Surveys 56.12 (Dec. 2024), pp. 1–37. issn : 0360-0300, 1557-7341. doi : 10.1145 / 3672392. url : (visited on 11 / 04 / 2024).

7] Qinyu Zhu et al. “Investigating TrustZone : A Comprehensive Analysis”. en. In : Security and Communication Networks 2023.1 , p. 7369634. issn : 1939-0122. doi : 10.1155 / 2023 / 7369634. url : (visited on 11 / 03 / 2024).

The PhD student will conduct original researches on the topic described above, and will collaborate with colleagues in the STACK team and Hive partners.

Activities includes, but are not limited to : bibliographical synthesis, research, proof writing, software implementation, presentation of results at conferences, attending research schools, etc

Technical skills and level required :

  • Solid understanding of cybersecurity principles such as threat modeling, cryptographic primitives and key management (PKI, symmetric / asymmetric encryption, hashing) and their integration in applications
  • Proficiency in systems programming (C / C++, Rust) and high level language (Python).
  • Experience with AI frameworks (PyTorch, Tensorflow, ...) and software performance measurement is a plus

Languages :

  • Good communication skills in English (French is a plus)

Relational skills :

  • Ability to work collaboratively in an academic–industry setting

Avantages

  • Partial reimbursement of public transport costs
  • Leave : 7 weeks of annual leave + 10 extra days off due to RTT (statutory reduction in working hours) + possibility of exceptional leave (sick children, moving home, etc.)
  • Possibility of teleworking (after 6 months of employment) and flexible organization of working hours
  • Professional equipment available (videoconferencing, loan of computer equipment, etc.)
  • Social, cultural and sports events and activities
  • Access to vocational training
  • Social security coverage

Salary gross : 2200€

Obtenez votre examen gratuit et confidentiel de votre CV.
ou faites glisser et déposez un fichier PDF, DOC, DOCX, ODT ou PAGES jusqu’à 5 Mo.