Welcome

Research Talks | Detection of Intrusions and Abnormal Behavior in Cybersecurity

We are pleased to invite you to the next Research Talks, a series of events designed to foster exchanges between academia and BNP Paribas experts around emerging technologies, research breakthroughs and their applications to the financial industry.

Organized within the framework of FinAI-LAB, this edition will explore some of the emerging challenges in federated learning and agentic AI, examining adversarial robustness, organizational patterns, and the evolving trade-offs between safety, security, innovation, and technological progress.

Date: Monday, 23 November 2026
Time: 10:00 – 12:00
Location: Grands Moulins de Pantin, 9 rue du Débarcadère / Bâtiment Afrique, Pantin

Featured Talks

Averaging Fails: Robust Aggregation in Federated Learning

Julien FAGEOT, Research Engineer, LTCI, Télécom Paris

In federated learning, workers send local updates, typically gradients, to a server that aggregates them. Plain averaging is the default choice, yet a single malicious worker can steer it arbitrarily far, degrading training or implanting backdoors. We present a framework for robust aggregation, built on a worst-case robustness coefficient that quantifies how much damage an omniscient adversary can inflict on an aggregator. We then refine this view by modeling concrete attacks and defining attack-specific robustness, which bridges the gap between worst-case guarantees and attacks observed in practice. Finally, we discuss ongoing directions that exploit the structure of the model, such as the layers of a neural network or LLM, viewing robustness as an attack–defense game.

Team Topologies in the Age of AI: Balancing Liveness and Safety

Peva BLANCHARD, Researcher, Kleis Technology

Every organisation trades off liveness (shipping, making progress) against safety (nothing bad happens). In banking, security tips the balance towards control. Drawing on field experience – hybrid on-premise and cloud services, product and platform teams around a software factory – I will show how some organisational patterns encode that trade-off. Agentic AI upends these balances and our preconceptions: the patterns have yet to be invented, and I will suggest a few avenues for sketching them.

Program

• 09:45 | Arrivals
• 10:00 | Introduction
• 10:10 | Talk by Julien FAGEOT
• 11:00 | Talk by Peva BLANCHARD
• 11:50 | Closing Remarks

FinAI-LAB is a joint Artificial Intelligence laboratory established by BNP Paribas and Télécom Paris (Institut Polytechnique de Paris). The initiative brings together researchers, research engineers, postdoctoral researchers, doctoral students and BNP Paribas experts to address strategic and high-impact challenges in Artificial Intelligence and Cybersecurity.