The defense will take place on March 02, 2026

Title
Service Provider Network Anomaly Detection: Standards, Architecture, and Algorithms for actual use by Network Operators

Abstract
IP networks are critical infrastructure. Disruptions can impact essential services and cause serious economic and societal harm. Rapid detection is therefore essential for service providers.
This thesis investigates an architecture and approach for detecting service anomalies in real-world service provider networks. While statistical outlier detection has been widely studied for this purpose, such methods often misalign with how engineers monitor and diagnose networks, limiting their actual adoption in production.
This thesis proposes a network-centric anomaly detection approach based on deterministic, rule-based inspections. Instead of outlier detection, expert knowledge is encoded into rules that reflect how engineers interpret network telemetry data. The system processes telemetry data, maps it to customer identifiers, and applies rules to detect deviations from expected behaviors. Because these rules mirror operational data inspections from network engineers, alerts are directly actionable: they identify affected customers and specify the underlying symptoms. We apply this method to the two most revenue-generating connectivity services: BGP/MPLS Layer 3 VPNs and Internet access.

Our work addresses three questions:
(i) what operational requirements must an anomaly detection system meet in service provider networks,
(ii) which telemetry protocols best align with network operations, and
(iii) which detection strategies best reflect engineers’ mental models.

This is the first systematic study of anomaly detection in BGP/MPLS Layer 3 VPNs. For Internet access, we focus on issues within the service provider network, including those invisible to end users but economically relevant. Our system is deployed in Swisscom’s production environment, monitoring over 13,000 VPN customers. It processes 760,000 telemetry messages per second and has detected incidents such as software defects, misconfigured DNS blacklists, and mistakenly decommissioned fiber links. Since its deployment, it has identified over 20 disruptions, enabling operators to quickly acknowledge incidents and postpone maintenance windows. This thesis presents production cases showing how the system reduced detection time and supported operational workflows.
Our findings show that effective anomaly detection in service provider environments must go beyond data-driven approaches. Systems need to reflect how network engineers actually work.Rule-based data inspections, grounded in established operational practices, offer a practical, actionable alternative to statistical models.

Jury
– CONTRERAS MURILLO Luis Miguel, Chercheur, Telefonica, Rapporteur
– PELSSER Cristel,  Professeur des Universités, UCLouvain, Rapporteuse
– IANNONE Luigi,  Chercheur, Huawei, Examinateur
– ROOSE Philippe, Professeur des Universités, IUT de Bayonne, Examinateur
– FRANCOIS Pierre, Professeur, INSA de Lyon, Directeur de thèse
– FRENOT Stéphane, Professeur des Universités, INSA de Lyon, Co-directeur de thèse

The defense will take place on wednesday 17th of september.

Title
Hardware Arithmetic Acceleration for Machine Learning and Scientic Computing.

Abstract
In a data-driven world, machine learning and scientific computing have become increasingly important, justifying dedicated hardware accelerators. This thesis explores the design and implementation of arithmetic units for such accelerators in Kalray’s Massively Parallel Processor Array.

Machine learning requires matrix multiplications that operate on very small number formats. In this context, this thesis studies the implementation of mixed-precision dot-product-and-add for various 8-bit and 16-bit formats (FP8, INT8, Posit8, FP16, BF16), using variants of a classic state-of-the-art technique, the long accumulator. It also introduces techniques to combine various input formats. Radically different methods are studied to scale to the larger range of 32-bit and 64-bit formats common in scientific computing.

This thesis also studies the evaluation of some elementary functions. An operator for exponential function (crucial for softmax computations) extends a state-of-the-art architecture to accept multiple input formats. The inverse square root function (used for layer normalisation) is accelerated by combining state-of-the-art techniques for range reduction, correctly rounded multipartite tables, and software iterative refinement techniques.

Date: 24/07/2025
Time: 10 am
Room: Amphitheatre Claude Chappe
Venue: Batiment Hedy Lamarr

Title
A study of energy consumption challenges in extended reality services over cellular networks

Abstract
Extended reality (XR) services are characterised by their heavy computational requirements throughout their life cycle. XR comprises multiple traffic modes consisting of 3D video and audio, haptics, sensor and pose information. Systems-related challenges relating to the creation, encoding, transmission, rendering and presentation of the application data from this class of services have increasingly become key areas of research inquiry from both a computational and energy viewpoint. Recently, new provisions have been made across different relevant standards to improve the capacity of these applications on mobile cellular networks. Our research investigates the energy-related challenges at both the RAN and UE levels.  We conduct our evaluations using system-level simulations (SLS) that adhere to the parameter specifications established by standardisation bodies.
We begin with a RAN-level energy reduction plus XR application enhancement strategy. With the cell switch off (CSO) technique proposed for BS deployment energy efficiency at low load, we evaluate the impact of an XR-capacity improvement criterion on the energy savings obtainable. Our analysis covers a European urban city using data from a European operator and system-level simulations according to the standard network operations.
In XR, UE energy-saving schemes are especially important since current delivery devices are still in early commercial development, making battery-saving techniques crucial. In this thesis, we use the Rel-18 improvements on the discontinuous reception (DRX) UE energy saving mechanism through state-of-the-art prediction algorithms to propose a traffic prediction-based non-integer DRX mechanism. We achieved significant energy savings without impeding the XR quality of service requirements.
Lastly, towards enhancing the capacity of XR on cellular networks in energy-saving mode, we propose a DRX plus QoE-aware scheduler (DQAS). Using SLS according to 3GPP specifications on XR QoE requirements, traffic model, and dense urban deployment scenarios, and following real traces from XR applications, we first evaluate a QoE-aware scheduler (QAS) for XR services. We introduced DRX awareness into QAS, jointly improving XR QoE and energy consumption. We analysed our results to identify the parameter window in which both metrics can be improved towards achieving the goal of XR capacity improvement on cellular networks.

Jury
– Riadh Dhaou, Professor des Universites (University of Toulouse), Rapporteur
– Thi-Mai-Trang Nguyen, Professeur des Universités (Sorbonne Paris Nord University), Rapporteure
– Nadjib Achir, Maître de Conférences HDR(Sorbonne Paris Nord University), Examinateur
– Herve Rivano, Professeur des Universités (INSA Lyon), Examinateur
– Razvan Stanica, Maître de Conférences HDR (INSA Lyon), Directeur de thèse