On January 28th, 2025, our research team attended FIIA 2025, the 9th industrial forum of the Association Française pour l’Intelligence Artificielle (AFIA). We presented the work conducted within the "Trustworthy and Responsible AI" research chair, launched in 2024 as a collaboration between Crédit Agricole and École Polytechnique. This chair aims to develop secure, ethical, and robust AI systems. Supported by engineers from Crédit Agricole's DataLab, it brings together a multidisciplinary team of researchers working on various topics, including AI explainability, fraud detection, robustness to attacks, model efficiency, and resource frugality.
As part of this research chair, we conducted studies to understand and mitigate AI-related privacy risks. Specialized attacks can extract confidential data used during model training, posing a particularly critical threat in the banking sector due to the sensitivity of the data involved. However, existing tools for measuring this risk and assessing model vulnerability are often complex and computationally expensive. To address this challenge, we developed a cost-effective approach to identify the most vulnerable training data, enabling an assessment of the model's privacy risk. Our method detects these data points before they are memorized by the model, facilitating the implementation of targeted mitigation techniques. This work is part of a broader effort to study adversarial attacks, understand the factors that make models vulnerable, and develop strategies to enhance their security. Our preprint is available here.
