Canada CIFAR AI Chair

Mathias Lécuyer

Academic Affiliations

University of British Columbia

Industry & Research Affiliations

Head of Research at Wiremind

Focus

Trustworthy AI, Privacy, Robustness, Causal Inference

Building Trustworthy AI

On topics ranging from privacy, robustness, explainability, and causality, with a specific focus on applications that provide rigorous guarantees.

Mathias works on trustworthy AI, on topics ranging from privacy, robustness, explainability, and causality, with a specific focus on applications that provide rigorous guarantees. Recent impactful contributions include: the first scalable defense against adversarial examples (small changes to inputs that can control AI models predictions, and be used for AI jailbreaks) with provable guarantees; a technique to efficiently measure the influence of training data on AI model behavior; a method to audit privacy leakage from AI models given only API access; and a system to enable federated and privacy preserving measurements of advertising performance, now serving as the blueprint for a future standard aiming to reduce third party tracking on the web.

His work has been recognized through spotlight and oral presentations at top conferences, a Google Scholar Award, and his appointment to the Research Council of the Canadian AI Safety Institute (CAISI) at CIFAR.

Three of his lines of research have had particularly high impact:

  • Certified Robustness: He designed the first, and still dominant, scalable defense against adversarial examples (small changes to inputs that can control AI models predictions, and be used for AI jailbreaks) with provable guarantees, helping develop the field of certified robustness and create the research area of randomized smoothing.

  • Privacy-Preserving Standards: He developed systems to enable federated and privacy preserving measurements of advertising performance, which now serve as the blueprint for a W3C standardization effort which, if successful, will reach hundreds of millions of user devices.

  • Causal Forecasting: His work on causal time-series forecasting is deployed to price tickets at several airlines and train companies (including in Canada), optimizing pricing decisions on billion dollars of revenue.

  • Kazmi, M., Lautraite, H., Akbari, A., Tang, Q., Soroco, M., Wang, T., ... & Lécuyer, M. (2024). Panoramia: Privacy auditing of machine learning models without retraining. Advances in Neural Information Processing Systems, 37, 57262-57300.

  • Lyu, S., Shaikh, S., Shpilevskiy, F., Shelhamer, E., & Lécuyer, M. (2024). Adaptive randomized smoothing: Certified adversarial robustness for multi-step defences. Advances in Neural Information Processing Systems, 37, 134043-134074.

  • Tholoniat, P., Kostopoulou, K., McNeely, P., Sodhi, P. S., Varanasi, A., Case, B., ... & Lécuyer, M. (2024, November). Cookie monster: Efficient on-device budgeting for differentially-private ad-measurement systems. In Proceedings of the ACM SIGOPS 30th Symposium on Operating Systems Principles (pp. 693-708).

  • Lin, J., Zhang, A., Lécuyer, M., Li, J., Panda, A., & Sen, S. (2022, June). Measuring the effect of training data on deep learning predictions via randomized experiments. In International Conference on Machine Learning (pp. 13468-13504).

  • Lecuyer, M., Atlidakis, V., Geambasu, R., Hsu, D., & Jana, S. (2019). Certified Robustness to Adversarial Examples with Differential Privacy. In 2019 IEEE Symposium on Security and Privacy (SP) (pp. 656-672).