Author: Krystyna Biletska / CEA
On September 17th 2025, CEA and Sorbonne University organized the Adversarial Threats on Real Life Learning Systems workshop in Paris. It focused on adversarial and backdoor attacks targeting real-life machine learning systems. The participants explored vulnerabilities in deployed learning systems, examine attack vectors in practical scenarios, and discuss defense mechanisms for robust ML deployment.
The workshop was inspired by research from the KINAITICS, which investigates kinematic indicators for adversarial behavior detection in AI systems. The event brought together researchers, academics, and industry professionals to discuss cutting-edge developments in adversarial machine learning, security implications, and mitigation strategies for production environments.
Among 33 participants: 2 keynote speakers, 7 speakers. Among 9 presentations: 6 scientific articles.
What stood out:
- Focus on real-world adversarial & backdoor attacks, not just theory
- Case studies including defect detection in physical systems, federated learning with zero-knowledge proofs, and interpreting neural network behaviour via topology
- Two keynotes: Kassem Kallas on backdoors (“Stealth Weapon or Structural Weakness?”) and Benjamin Negrevergne on adversarial attacks & mitigation strategies
The program and several articles remain available on the webpage.