Dyman was present at the Thirteenth International Conference on Learning Representations (ICLR) held in Singapore from 24 to 28th April. The conference is aimed at the scientific community and industry stakeholders with and estimated attendance of 10000 people during the five-day event.
ICLR is a premier venue for machine learning and a top AI conference with high visibility. It is primarily a scientific conference featuring poster sessions. Our team achieved the spotlight badge (given to 3.26% of the submitted papers) and presented our work in one of the poster sessions. One hall had a space reserved for about 20 stands from different companies with an interest in AI, ranging from Big Tech (e.g., Google and Meta) to Robotics and Trading.
The work is a collaboration with the University of Neuchâtel, TU Delft, and IBM Research. The work presented focuses on watermarking synthetic tabular data generated by diffusion models. We embed watermarks into the latent noise used as input for the diffusion model rather than the synthetic adata itself. Hence, this method can achieve close to no impact on the perceived quality while still allowing to track the provenance of the generated data.
We are looking forward to present further advances in the project.
