DeepMark, a start-up company working on development of invisible and permanent digital marks (watermarks) that can be embedded in images, audio, and video content without compromising quality has made successful application to EuroHPC JU Development call, thus gaining 4500 node hours to Leonardo Booster partition for a period of one year. NCC Montenegro supported this project application.
Provenance for synthetic media remains unsolved: most watermarks fail after routine processing, and especially under modern AI transforms. DeepMark advance a learning-based system that co-trains a neural embedder and detector through a differentiable signal chain that mimics real pipelines. Goals of the project are: (i) lower bit-error rates at equal perceptual quality; (ii) patchable watermarking so future embedders remain detectable by a single detector; and (iii) calibrated, segment-level detection for forensics and compliance. Access to A100 nodes and a parallel file system is required to tune multi-GPU scaling and remove I/O bottlenecks.