INGI lunch time seminar
ingi |
INGI Lunchtime Seminars take place about twice a month and they address various topics related to computing sciences.
They are open to everyone, no registration is required : come and have a seat! Sandwiches and drinks are usually provided. Please fill in the form in the invite before day D-1 at 12:00 to reserve a sandwich
If you are not an ICTEAM member, join our mailing list to stay informed about the next seminars!
INGI Lunch time seminars
Seminars to Come
[INGI] 2026-07-02 (13:00) : 2 INGI seminars in one (J. Vanderhaeghen & A. Fierens)
At Shannon (Maxwell - a.105)
Speaker:
Amaury Fierens (ICTEAM)
,
and
Juliette Vanderhaeghen (ICTEAM)
Abstract: By Juliette Vanderhaeghen:
Towards Personalized Tumor Growth Modeling with Physics-Informed Neural Networks
Abstract:
Cancer remains a major global health challenge, highlighting the urgent need for more personalized treatment strategies. Mathematical models, in particular PDE-based approaches, provide a powerful framework to describe tumor dynamics. However, classical numerical methods can be computationally expensive and require careful calibration, while purely data-driven models often lack interpretability and demand large amounts of data.
Physics-Informed Neural Networks (PINNs) offer a promising alternative by embedding physical laws directly into the training process, enabling both efficient simulations and the inference of patient-specific parameters from limited data.
In this presentation, I present our current work on the use of PINNs to model tumor growth dynamics. First, I will demonstrate how PINNs can reconstruct tumor evolution and estimate key parameters from partial synthetic observations. Then I will present their application to multicellular tumor spheroids as a controlled experimental system. These spheroids generate spatially and temporally resolved data under varying treatment conditions, making them a valuable testbed for model validation.
By Amaury Fierens:
Self-Alignment and Reranking for Large-Scale Medical Coding in Open-Source Clinical NLP
Abstract:
Medical coding is a challenging large-scale linking problem: clinical mentions must be mapped to highly ambiguous terminologies containing millions of possible codes, such as UMLS. While large language models are increasingly used in clinical NLP, they are not yet a fully satisfactory solution for this task, especially when reliability, traceability, terminology coverage, and precise code selection are required.
The presentation is structured in three parts. First, I will introduce the complexity of medical coding and why it remains difficult for both automated systems and clinicians. Second, I qill discuss an effective technical approach based on self-alignment pretraining with contrastive learning, as in SapBERT, combined with candidate reranking through cross-encoders, rule-based signals, and hierarchical information. Finally, drawing on the open-source work carried out this year by two master’s students, I will discuss how effective NLP components can be integrated into practical tools that support medical annotation and coding workflows.
Towards Personalized Tumor Growth Modeling with Physics-Informed Neural Networks
Abstract:
Cancer remains a major global health challenge, highlighting the urgent need for more personalized treatment strategies. Mathematical models, in particular PDE-based approaches, provide a powerful framework to describe tumor dynamics. However, classical numerical methods can be computationally expensive and require careful calibration, while purely data-driven models often lack interpretability and demand large amounts of data.
Physics-Informed Neural Networks (PINNs) offer a promising alternative by embedding physical laws directly into the training process, enabling both efficient simulations and the inference of patient-specific parameters from limited data.
In this presentation, I present our current work on the use of PINNs to model tumor growth dynamics. First, I will demonstrate how PINNs can reconstruct tumor evolution and estimate key parameters from partial synthetic observations. Then I will present their application to multicellular tumor spheroids as a controlled experimental system. These spheroids generate spatially and temporally resolved data under varying treatment conditions, making them a valuable testbed for model validation.
By Amaury Fierens:
Self-Alignment and Reranking for Large-Scale Medical Coding in Open-Source Clinical NLP
Abstract:
Medical coding is a challenging large-scale linking problem: clinical mentions must be mapped to highly ambiguous terminologies containing millions of possible codes, such as UMLS. While large language models are increasingly used in clinical NLP, they are not yet a fully satisfactory solution for this task, especially when reliability, traceability, terminology coverage, and precise code selection are required.
The presentation is structured in three parts. First, I will introduce the complexity of medical coding and why it remains difficult for both automated systems and clinicians. Second, I qill discuss an effective technical approach based on self-alignment pretraining with contrastive learning, as in SapBERT, combined with candidate reranking through cross-encoders, rule-based signals, and hierarchical information. Finally, drawing on the open-source work carried out this year by two master’s students, I will discuss how effective NLP components can be integrated into practical tools that support medical annotation and coding workflows.
Previous Seminars
Page 1 of 1
[INGI] 2026-06-04 (13:00) : Human-like game playing beyond behavioral alignment
At Shannon (Maxwell - a.105)
Speaker:
Aloïs Rautureau (ENS Rennes)
Abstract: Game playing systems have been extensively studied with the prospect of improving their playing strength, but little work has been done on making them play "like humans". Such "human-like" systems would have a number of advantages and potential use cases: designers wanting to test early prototypes of their game quickly and extensively, players hoping to learn actionable and realistic strategies, or researchers in humanities trying to better understand how human players approach ancient games. While current approaches center their definition around the Turing test (an agent is human-like if it is indistinguishable from one), and are thus mainly based around imitation learning, we argue that human-like play is broader than move selection. The way humans approach games is deeply rooted in cognition, psychology, and the social context in which they play. This seminar will present exploratory work on defining what "human-like" even means in the context of game playing agents, present methods and algorithms aiming to bridge the gap between human cognition and computational approaches, and discuss future work and research directions that stem from this exploration.
[INGI] 2026-05-13 (13:00) : Scala, WebAssembly, and JavaScript interoperability
At Shannon (Maxwell - a.105)
Speaker:
Sébastien Doeraene (EPFL, Switzerland)
Abstract: Since Scala.js 1.17.0, released in September 2024, users have been able to target WebAssembly (Wasm) for their applications. With a single build tool configuration switch, the Scala.js linker generates Wasm code to be used in a JavaScript host. It does so while preserving the semantics of the Scala.js language. In particular, you still get all the nice JavaScript interoperability features that Scala.js is known for (well … except @JSExport, but more on that in the talk). That also means you get to reuse the entire ecosystem of libraries. This is unusual among languages targeting Wasm. Most significantly compromise on their ability to talk to JavaScript in the process.
This talk explores some of the challenges we faced, and the unusual solutions we designed. We will talk about weird language semantics, low-level performance aspects (including where JS still beats Wasm), and even new proposals to Wasm.
[INGI] 2026-05-12 (13:00) : AI for societal impact: research at the University of Abomey-Calavi, Benin
At Shannon (Maxwell - a.105)
Speaker:
Ratheil V. Houndji (UAC Benin)
Abstract: This presentation showcases AI research conducted at the University of Abomey-Calavi (UAC) that addresses real-world challenges in agriculture, health, and optimization. It begins with an overview of the AI ecosystem at the UAC, followed by a focus on our team's work in the “Laboratoire de Recherche en Sciences Informatiques et Applications (LRSIA)”. The talk presents selected projects and ongoing work in three main areas:
- AI for Agriculture, including research on:
- Optimizing tomato yield using machine learning techniques
- Modeling maize (Zea mays) yield in Benin
- Monitoring cotton jassid infestation
- AI for Health, including research on:
- AI for Chronic Kidney Disease
- AI for neurological diseases, specifically for epilepsy
- Optimization, including research on:
- Bias evaluation and mitigation in datasets using constraint programming
- VRP-CC
[INGI] 2026-04-16 (13:00) : No evaluation without fair representation: Impact of label and selection bias on the evaluation, performance and mitigation of classification models
At BARB 94 (Hall Sainte-Barbe)
Speaker:
Magali Legast (ICTEAM)
Abstract: Bias can be introduced in diverse ways in machine learning datasets, for example via selection or label bias. Although these bias types in themselves have an influence on important aspects of fair machine learning, their different impact has been understudied. In our work, we empirically analyze the effect of label bias and several subtypes of selection bias on the evaluation of classification models, on their performance, and on the effectiveness of bias mitigation methods. We also introduce a biasing and evaluation framework that allows to model fair worlds and their biased counterparts through the introduction of controlled bias in real-life datasets with low discrimination. Using our framework, we empirically analyze the impact of each bias type independently, while obtaining a more representative evaluation of models and mitigation methods than with the traditional use of a subset of biased data as test set. Our results highlight different factors that influence how impactful bias is on model performance. They also show an absence of trade-off between fairness and accuracy, and between individual and group fairness, when models are evaluated on a test set that does not exhibit unwanted bias. They furthermore indicate that the performance of bias mitigation methods is influenced by the type of bias present in the data. Our findings call for future work to develop more accurate evaluations of prediction models and fairness interventions, but also to better understand other types of bias, more complex scenarios involving the combination of different bias types, and other factors that impact the efficiency of the mitigation methods, such as dataset characteristics.
[INGI] 2026-03-24 (11:00) : 2 INGI seminars in one
At Shannon (Maxwell - a.105)
Speaker:
Martin Henze (RWTH Aachen University)
,
and
Anna Maria Mandalari (University College London)
Abstract: "Strengthening the IoT Ecosystem: Privacy Preserving IoT Security Management"
by Anna Maria Mandalari (University College London)
Abstract:
As the Internet of Things (IoT) continues to proliferate in our daily lives, concerns about the potential intrusion into our privacy and data security have become more prevalent. This talk aims at fostering an IoT ecosystem that prioritizes user privacy, security, efficiency, and reliability. In an age where smart city initiatives and connected homes are becoming the norm, it is crucial to address the growing challenges associated with IoT devices. This talk will explore strategies and solutions for establishing an IoT environment that places user interests at the forefront. The discussion will encompass cutting-edge approaches to safeguarding data, preserving privacy, and enhancing security in IoT networks, offering a vision of a future where our connected world is safer, more efficient, and truly user-centered.
------------------------
"Generalizable and Comprehensible Intrusion Detection for Industrial Control Systems"
by Martin Henze (RWTH Aachen University)
Abstract:
Securing industrial control systems against cyberattacks is crucial to
counter imminent threats to critical infrastructure. Intrusion detection
provides an easily retrofittable approach to timely uncover attacks
before they can cause substantial damage. However, research on intrusion
detection for industrial control systems suffers from the use of complex
approaches hindering the interpretability of alarms as well as
non-expressive evaluations, often bound to one system with unknown
generalizable to other deployments. This talk presents our research
platform for deployment-independent intrusion detection as well as
approaches towards comprehensible alarms, thus making research
advancements in intrusion detection more accessible to industrial
control system deployments.
[INGI] 2026-03-12 (13:00) : When AI co-designs and implements a programming language : the case of Elo
At Shannon, Maxwell a.105
Speaker:
Bernard Lambeau (Klaro Cards)
Abstract: I launched Claude Code on the design and implementation of the Elo data expression language the 24th of December. By the 1st of January, the compiler, website and documentation were ready for pre-production. I did not touch a single line of code, yet a week later I was confident enough embedding Elo in production systems. In this talk I'll share my motivation for creating Elo, the methodology I used with Claude Code, and lessons learned about integrating AI in Software Engineering processes in my companies.
[INGI] 2026-03-09 (13:00) : Development of digital twins of reference softwares for testing/certification in realistic environment.
At BARB 94
Speaker:
Ahmed Bokri (ERM)
Abstract: Security certification of detection systems requires testing environments that are realistic but also fully controlled. This presentation describes the development of a digital twin of the Multi-Agent System for APT Detection (MASFAD), deployed on the KYPO Cyber Range to support structured testing and certification activities.
The digital twin reproduces the software architecture of MASFAD and the environment in which it operates. It includes a realistic enterprise-like infrastructure, background traffic generation, and controlled APT-inspired persistence scenarios. All system and security events are centrally collected through the Elastic Stack to ensure clear visibility of system activity and detection results.
The platform ensures reproducibility, traceability, and controlled evidence generation. Experiments can be repeated under the same conditions, and the results can be verified. This work shows how a digital twin can be used as a practical and structured environment for evaluating detection capabilities in view of formal security certification.
[INGI] 2026-03-05 (13:00) : xPUBench: Scalable and Energy-Efficient GPU and DPU-Accelerated Network Functions
At Nyquist Maxwell a.164
Speaker:
Maxime Vanliefde (ICTEAM)
Abstract: The rapid increase in network speeds makes packet processing on general-purpose CPUs increasingly challenging. At 100 Gbps and beyond, CPUs struggle to sustain complex network functions without dedicated acceleration. This trend motivates the exploration and measurement of alternative compute platforms such as GPUs and embedded CPUs in Network Interface Cards (NICs). Modern NICs provide tighter integration with GPUs, with the ability to write received packets directly to GPU memory. SmartNICs, also known as DPUs, further feature embedded ARM or RISC cores capable of offloading NFV packet processing entirely. In this work, we introduce xPUBench, a benchmarking environment that systematically measures the performance and energy efficiency of packet processing across CPUs, GPUs, and DPUs. We evaluate several (co-)processing models relevant to Network Function Virtualization, including CPU+GPU hybrid, DPU-only, and GPU-only approaches. Our measurements show that, for a computation-heavy workload, current CPU-only implementations manage to handle up to 50% of the 100 Gbps NIC rate. In contrast, GPU implementations can saturate it. We also show that SmartNICs’ most powerful embedded cores can replace the main CPU for some traditional packet processing, alleviating the load on the host, which can now be entirely dedicated to running applications. We finally propose a novel energy-efficiency dimension, showing that DPUs outperform traditional CPUs for low-throughput processing, requiring only 24 W to sustain 10 Gbps, and that GPUs outperform CPUs for high-throughput processing. Our findings emphasize the need to assess both performance and energy in heterogeneous packet-processing pipelines, given the growing diversity of “xPUs” in networked systems.
[INGI] 2026-02-25 (11:00) : Closed Loop Automation System for xG networks
At Shannon
Speaker:
Stefano Secci (CNAM, France)
,
and
Patient NTUMBA WA NTUMBA (CNAM Paris)
Abstract: In this presentation, we present a closed-loop automation system that leverages in-network distributed learning to automatically mitigate anomalous states in the connect-compute software infrastructure. We describe the key components of the automation system, including: an anomaly detection module based on federated learning; AI function scheduling to meet detection performance targets while mitigating federated learning stragglers; a data-pipeline design that supports high accuracy, real-time preprocessing, and timely data delivery; a network data sources load-balancing strategy for federated-learning clients; and automated reconfiguration management using deep reinforcement learning (DRL). We also deliver a live demo of the core system blocks—showing, in real time, the anomalies detected during the inference phase using federated learning with real-time data pipelining.
[INGI] 2026-02-19 (13:00) : Why AI is eating the planet.
At BARB94
Speaker:
Romain Rouvoy (Full Professor within the Faculty of Science and Technology at the University of Lille and the head of the Computer Science department)
Abstract: This presentation aims to raise awareness of the environmental costs of our personal and professional activities, which are increasingly assisted by AI agents. In particular, throughout this presentation, we aim to provide greater transparency into the conditions under which these tools are deployed, enabling a more accurate estimate of the consumption attributable to their use. To guide this approach, we study the production of a line of source code as a functional unit, enabling us to approximate the cost of an AI assistant's impact across the many infrastructure layers involved in its usage phase.
[INGI] 2026-02-12 (10:45) : Data Analytics : Data-driven consumer centric services
At BARB 94
Speaker:
Christophe Robyns (Agilytic)
Abstract: Massive internal and external data are available to better understand customer sentiment and anticipate or influence his behaviour. Having experience both in commercial and non-commercial organisations, the speaker will explain how data can be used to influence customer behaviour and enhance services. The speaker will also present steps that are required to build a big data platform that provides reliable and up-to-date information.
[INGI] 2026-02-05 (13:00) : Fair Tabular Data Generation: an Approach using Autoregressive Decision Trees
At Nyquist Maxwell a.164
Speaker:
Benoît Ronval (ICTEAM)
Abstract: In both research and industry, tabular data is among the most widely used data types. Represented using instances (rows) with features (columns), such data is easy for humans to interpret and readily usable by most machine learning algorithms.
Despite its large usage, acquiring new tabular data can be challenging. Data collection can require access to private sources or large-scale surveys, which may be costly and may suffer from low response rates. Moreover, real-world tabular datasets frequently exhibit bias, leading machine learning models to produce unfair classifications for certain subgroups, particularly with respect to sensitive attributes such as the nationality or the education level of a person.
In this seminar, I will present our new method TabFairGDT, which aims to generate data that can reduce fairness concerns in the predictions of machine learning models trained on this data. The approach leverages decision trees in an autoregressive generation framework, including a fairness optimization step. I will also discuss the advantages of decision trees for tabular data generation and present experimental results, including classification performance, fairness metrics, and data quality analyses.
[INGI] 2026-01-23 (11:00) : Neuro-symbolic Deep Learning with Requirements
At Shannon, Maxwell a.105
Speaker:
Eleonora Giunchiglia (Assistant Professor in Machine Learning and AI at Imperial College London | PI of DUCK Lab)
Abstract: For their outstanding ability of finding hidden patterns in data, deep learning models have been extensively applied in many different domains. However, recent works have shown that, if a set of requirements expressing inherent knowledge about the problem at hand is given, then neural networks often fail to comply with them. This represents a major drawback for deep learning models, as requirements compliance is normally considered a necessary condition for standard software deployment. We present a neuro-symbolic framework able to make any neural network compliant by design to a given set of requirements over the output space expressed in full propositional logic. This framework integrates the requirements into the output layer of the neural network.
[INGI] 2025-12-11 (13:00) : Combinatorial Transition Testing in Dynamically Adaptive Systems: Implementation and Test Oracle
At Shannon, Maxwell a.105
Speaker:
Pierre Martou (ICTEAM)
Abstract: Due to the large number of possible interactions and transitions among features in dynamically adaptive systems, testing such systems poses significant challenges. To verify that such systems behave correctly, combinatorial interaction testing (CIT) can create concise test suites covering all valid pairs of features of such systems. While CIT claims to find all errors caused by two features, it does not cover certain errors occurring only for specific transitions between features. To address this issue we study the technique of Combinatorial Transition Testing (CTT), which includes both generation and detection of what we call behavioural transition errors. From an initial generation algorithm that combines both interaction and transition coverage but lacks scalability, we propose an optimised version that enables CTT even for hundreds of features. From a valid test suite covering all transitions, we complete our testing approach with a test oracle that detects all behavioural transition errors without any prior knowledge of the system's behaviour. After a comprehensive analysis over a large number of feature models, we conclude that size of CTT-generated test suites and test effort needed to use our test oracle are linearly correlated to CIT-generated ones and that CTT grows logarithmically in the number of features.
[INGI] 2025-12-04 (13:00) : Empowering QUIC with Connection Delegation to Save Energy
At Shannon, Maxwell a.105
Speaker:
Vany Ingenzi (ICTEAM)
Abstract: The end-to-end principle is fundamental to the design of the Internet, requiring that hosts participating in a transport connection remain continuously capable of processing packets. However, this principle conflicts with the energy-saving needs of battery-powered wireless devices, which benefit from short and intermittent activity periods.
In this talk, we present a QUIC connection delegation mechanism that allows a client to transfer an active connection to a trusted third party by securely migrating its connection state. The third party then maintains the connection and handles data exchanges on behalf of the original client, allowing the client device to enter a power-saving mode. Our evaluation shows that QUIC connection delegation can reduce energy consumption over Wi-Fi by up to 55% when the smartphone can utilize its sleep mode, and by 25% when accounting for background traffic and an active display.
[INGI] 2025-11-27 (13:00) : Uncovering Malicious Persistence: Machine Learning-Based Detection of Windows Scheduled Tasks
At Shannon, Maxwell a.105
Speaker:
Khaled Rahal (ERM & ULB)
Abstract: Advanced Persistent Threats (APT) represent a serious security concern because they
carry out long-term and carefully planned attacks. While a lot of research has gone into finding
ways to detect these threats; one crucial area often gets less attention, namely the persistence
mechanisms that allow attackers to stay hidden and maintain access to systems over
time. In this work, we investigate scheduled tasks, a widely used persistence technique
in Windows environments, and analyze their role in APT operations. We conducted an
in-depth study of how attackers leverage scheduled tasks to maintain stealthy access and
execute malicious actions over time. We introduce Detecting APT Through Malicious
Scheduled Tasks (DAPTASK), an approach that leverages Sysmon log data, Word2Vecbased
feature representation, and Machine Learning (ML) classifiers to identify malicious
1scheduled tasks commonly used in APT persistence techniques. Our approach achieves
a high detection performance, with an F1-score of 95.19%. Furthermore, we provide a
labeled dataset, which can serve as a valuable resource for researchers developing APT
detection methods, the dataset and the code used are publicly available at https://gitlab.cylab.
be/cylab/daptask. Our approach enhances APT detection by addressing persistence
techniques, a critical yet often neglected attack vector.
[INGI] 2025-11-17 (13:00) : From Cloud to Edge: How to Make AI Inference Accurate, Fast and Resource-Aware ?
At Nyquist Maxwell a.164 room.
Speaker:
Patient NTUMBA WA NTUMBA (CNAM Paris)
Abstract: Edge computing is becoming a critical enabler for real-time Internet of Things (IoT) applications powered by artificial intelligence (AI). These applications often require accurate, low-latency and high-throughput inference while operating on resource-constrained edge servers. In this seminar, we present a holistic framework for optimising AI model provisioning at the edge. Our approach combines HProfiler, a data-driven model profiling tool that generates AI model variants tailored for accuracy-throughput trade-offs and specific edge resources, with AIForwarder, a reinforcement learning-based mechanism that dynamically manages model activation and request-to-model forwarding. By balancing accuracy, energy consumption, and request loss, our solution adapts to dynamic workloads and heterogeneous IoT demands.
Short Bio: Patient Ntumba is a postdoctoral researcher in the RoC team at the Conservatoire National des Arts et Métiers (CNAM) in Paris, France. He obtained his PhD from Sorbonne University, conducting his doctoral research at the INRIA Paris research center in the MiMove team. His research interests focus on networks, distributed systems, and optimization. His current work centers on in-network distributed learning and Edge AI inference.
Short Bio: Patient Ntumba is a postdoctoral researcher in the RoC team at the Conservatoire National des Arts et Métiers (CNAM) in Paris, France. He obtained his PhD from Sorbonne University, conducting his doctoral research at the INRIA Paris research center in the MiMove team. His research interests focus on networks, distributed systems, and optimization. His current work centers on in-network distributed learning and Edge AI inference.
[INGI] 2025-11-10 (11:00) : Double seminars by Hugo Rimlinger (LIP6) & Moritz Müller (université de Twente)
At Shannon room, Maxwell building
Section 1: GeoResolver: An Accurate, Scalable, and Explainable Geolocation Technique Using DNS Redirection
Speaker:
Hugo Rimlinger (LIP6)
Abstract: Obtaining an accurate, explainable and Internet scale IP geolocation dataset has been a longstanding goal of the research community. Despite decades of research on IP geolocation, no current technique can provide such a dataset. In particular, latency-based geolocation techniques do not scale, because, on one hand, we have thousands of available vantage points to perform measurements, but on the other hand, we have no way to select the right ones for each IP address. In this paper, we present GeoResolver, which is a serious step towards our goal, by using the idea that when multiple operators redirect two prefixes to the same servers, these prefixes should be close to each other. With this intuition, we define a methodology to measure and compare the redirection of prefixes to servers using ECS DNS measurements, and select the prefixes with the smallest redirection distance to a target prefix to issue the latency measurements to targets in that prefix. GeoResolver performs nearly as well as a brute force approach, geolocating 94% of the targets that could actually be geolocated at metro level, while using 4.3% of the probing budget compared to the state of the art. On the Internet scale CAIDA ITDK dataset, GeoResolver geolocates 16% of the IP addresses at metro level, 3.4 times more than the state of the art. In addition, GeoResolver is robust to public resolvers or hypergiants stopping supporting ECS.
Section 2: Monitoring highly distributed DNS deployments: challenges and recommendations for the root server system
Speaker:
Moritz Müller (Université de Twente)
Abstract: DNS name servers are crucial for the reachability of domain names. For this reason, name server operators rely on multiple name servers and often replicate and distribute each server across different locations across the world. Operators monitor the name servers to verify that they meet the expected performance requirements. Monitoring can be done from within the system, e.g. with metrics like CPU utilisation, and from the outside, mimicking the experience of the clients. In this talk, we focus on the latter. We take the root server system as a use case and highlight the challenges operators and researchers face when monitoring highly distributed DNS deployments from the outside. We also present recommendations on building a monitoring system that is more reliable and that captures only the relevant metrics.
[INGI] 2025-10-30 (13:00) : Hybrid artificial intelligence for solving computationally hard problems
At Shannon, Maxwell a.105
Speaker:
Quentin Cappart (ICTEAM)
Abstract: Combinatorial optimization provides methods to make the best possible decisions in complex scenarios, with practical applications in areas such as transportation, logistics, and healthcare. Traditionally, solving methods for combinatorial problems (such as integer programming, constraint programming, or local search) have focused on solving isolated problem instances, often overlooking the fact that these instances frequently originate from related data distributions. In recent years, there has been a growing interest in leveraging machine learning, particularly neural networks, to enhance combinatorial solvers by utilizing historical data. Despite this interest, it remains unclear how to effectively integrate learning into such engines to boost overall performance. In this presentation, I will share my journey in tackling this challenge, from my initial attempts to my current research directions. I will offer personal advice for researchers interested in exploring this fascinating field, highlighting the potential and opportunities for designing efficient hybrid artificial intelligence for solving computationally hard problems.
[INGI] 2025-10-16 (13:00) : Malware Detection with Machine Learning: Challenges and Perspectives
At Shannon, Maxwell a.105
Speaker:
Serena Lucca (ICTEAM)
,
and
Samy Bettaied (ICTEAM)
Abstract: This presentation explores some challenges and perspectives of applying machine learning to malware detection, drawing insights from two complementary studies. The first investigates the surprising effectiveness of simple models—such as One-Rule and AdaBoost—on state-of-the-art malware detection datasets, revealing that a small subset of dominant features often drives classification performance, leading to minimal differences between simple and deep learning approaches. The second study provides a systematic comparison of tabular and graph-based feature representations under unified conditions, evaluating their trade-offs in computational cost, detection accuracy, and robustness to adversarial attacks. Together, these works question the common assumption that complex models or sophisticated feature types always yield superior results, and instead highlight the importance of understanding feature dominance, dataset biases, and practical constraints in real-world malware detection.
[INGI] 2025-10-15 (10:00) : Autonomous Systems under AReST: Advanced Revelation of Segment Routing Tunnels
At Shannon, Maxwell a.105
Speaker:
Florian Dekinder (ULiège)
Abstract: Segment Routing (Sr), an advanced source routing mechanism, is a promising technology with a wide range of applications that has already gained traction from hardware vendors, network operators, and researchers alike. However, despite the abundance of activity surrounding Sr, little is known about how to gauge Sr deployment and its usage by operators.
This paper introduces a methodology, called AReST (Advanced Revelation of Segment Routing Tunnels), for revealing the presence of Sr with Mpls as forwarding plane (Sr-Mpls). AReST relies on standard measurement tools, like traceroute and fingerprinting, and post-processes the collected data for highlighting evidence of Sr-Mpls. Our results show that AReST is efficient in revealing the presence of Sr-Mpls in various autonomous systems, obtaining a perfect precision on our ground truth directly obtained from an operator. We also make a preliminary characterization of the Sr-Mpls deployment and show that it is commonly deployed within Content, Transit, and Tier-1 providers and, occasionally, in interworking with classic Mpls. The data collected, as well as our source code, is available to the research community.
[INGI] 2025-10-09 (13:00) : Sloth: A Kernel-Bypass Scheduler Maximizing Energy Efficiency under Latency Constraints
At Shannon room, Maxwell building
Speaker:
Clément Delzotti (ICTEAM/UCLouvain)
Abstract: The continuously increasing network speeds make packet processing on CPUs increasingly challenging. At a line rate of 100 Gbps, today’s CPUs struggle to execute complex network functions. This trend calls for offloading packet processing to other devices. This work explores how Graphical Processing Units and programmable Network Interface Cards can be used instead of a general-purpose CPU for packet processing. GPUs have been proposed to accelerate network processing thanks to their massively parallel architectures. Recent NICs provide tighter integration with GPUs, with the ability to write received packets directly to GPU memory. Recent SmartNICs, or DPUs, can also receive packets to their own memory and process them with embedded ARM or RISC processors. Such improvements allow bypassing the CPU entirely by processing packets only on the GPU cores or on the SmartNIC’s cores.
In this work, we review various models for packet (co-)processing applied to Network Function Virtualization, including CPU+GPU hybrid, SmartNIC-only, and GPU-only approaches. We introduce a novel communication model between
CPU cores and the GPU, enabling scalable CPU-GPU hybrid utilization while minimizing CPU resources needed. We show that for a computation-heavy workload, current CPU-only implementations manage to handle up to 45 % of the 100 Gbps line rate. In contrast, GPU implementations can saturate it. We also show that recent SmartNICs are getting powerful cores that can replace the main CPU for some traditional packet processing, alleviating the load on the host, which can now entirely be dedicated to running applications. We finally propose a novel energy-efficiency aspect, showing that GPUs and DPUs outperform traditional CPUs by 2 to 3× in terms of Joules/packet.
[INGI] 2025-10-09 (13:00) : Sloth: A Kernel-Bypass Scheduler Maximizing Energy Efficiency under Latency Constraints
At Shannon, Maxwell a.105
Speaker:
Clément Delzotti (ICTEAM/UCLouvain)
Abstract: In recent years, multi-hundred-gigabit networking
applications such as Virtual Network Function (VNF) and Key
Value Store (KVS) implementations have relied on kernel-bypass
and polling to achieve maximum throughput. However, this
performance improvement comes at the expense of high CPU
usage and power consumption.
This paper first analyses the trade-off between the power
consumption, the latency and the throughput of VNF applica-
tions. We then present Sloth, an energy-aware scheduler that
adapts the number of cores used by an application and their
frequency. Sloth uses the information gathered in a training phase
to maximize the energy reduction in real time while maintaining
a user-provided service-level objective. Sloth manages to reduce
CPU power consumption by up to 50% compared to the classical
DPDK polling approach with only a 30 μs latency increase. Sloth
also saves up to milliseconds of latency compared to state-of-the-
art solutions at equivalent power consumption
[INGI] 2025-09-25 (13:00) : Learning from logical constraints with Lower- and Upper bound arithmetic circuits
At Shannon, Maxwell a.105
Speaker:
Alexandre Dubray (ICTEAM)
Abstract: In this work focuses on the field of 𝐧𝐞𝐮𝐫𝐨-𝐬𝐲𝐦𝐛𝐨𝐥𝐢𝐜𝐥𝐞𝐚𝐫𝐧𝐢𝐧𝐠 (NeSy), which aims to bridge the gap between deep learning methods (neural) and the logical knowledge available in certain domains (symbolic). It has been accepted to the Main track at IJCAI 2025, one of the world’s premier conferences on Artificial Intelligence.
Standard deep learning struggles with logic-based reasoning. One solution is to encode known constraints as 𝐚𝐫𝐢𝐭𝐡𝐦𝐞𝐭𝐢𝐜 𝐜𝐢𝐫𝐜𝐮𝐢𝐭𝐬 to enable a gradient-based guidance of the parameters being learned. But, for logical knowledge that is too complex to be fully encoded, existing methods use a single lower-bound approximate circuit, often compromising the quality of the computed gradients.
The authors introduce a 𝐝𝐮𝐚𝐥-𝐛𝐨𝐮𝐧𝐝 𝐚𝐩𝐩𝐫𝐨𝐚𝐜𝐡, 𝐮𝐬𝐢𝐧𝐠 𝐛𝐨𝐭𝐡 𝐚 𝐥𝐨𝐰𝐞𝐫- 𝐚𝐧𝐝 𝐚𝐧 𝐮𝐩𝐩𝐞𝐫-𝐛𝐨𝐮𝐧𝐝 𝐜𝐢𝐫𝐜𝐮𝐢𝐭, to tightly control the error in the gradient approximation. This improves the robustness and trustworthiness of constraint-based learning.
[INGI] 2025-08-26 (11:00) : Robustness for Tabular Machine Learning: A Back-and-Forth Journey between Research and Industry
At Shannon room
Speaker:
Maxime Cordy (SnT Université du Luxembourg)
Abstract: Adversarial attacks are widely recognized as a critical security threat for machine learning, but most were originally designed in the context of image recognition, where arbitrary pixel-level perturbations are applied. Such attacks often fail to capture the realities of domains governed by strict constraints on valid inputs. This is particularly the case in tabular machine learning, where only feasible feature-level perturbations can occur. This talk explores the gap between classical adversarial attack formulations and real-world applicability. I will review our recent research on constrained feature-space attacks, which aim to generate realistic adversarial examples under domain-specific restrictions. Drawing on our own experience in applying these methods to industrial use cases, I will highlight the challenges of evaluating robustness in practice and discuss opportunities for new research in this area.
Page 1 of 1
Please visit seminars archive for the whole list.