Working on securing software against backdoors and hidden exploits using a set of debloating tools. First one available here: github.com/negativa-ai/BLAFS
The tool claims to reduce the GPU code size by up to 75% and the CPU code by up to 72% for ML workloads, resulting in total file size reductions of up to 55%.They also claim to achieve reductions in peak CPU memory usage, peak GPU memory usage, and execution time by up to 74.6%, 69.6%, and 44.6%. Authors claim they will opensource the tools here: https://github.com/negativa-ai
So we have been working on a solution to this problem for the past 5 years at university. We have just released one tool for containers (not the full thing for now) and we are about to release our tools for removing bloat in shared libraries. Out paper describing one of these tools won the best paper award at MLSys yesterday! https://mlsys.org/virtual/2025/poster/3238
If there are any adopted or anyone who would like to try our tools, please reach out! We would love to support you!
We are a bunch of academics who have worked on debloating tools for containers and we just released our code with an MIT license to Github: https://github.com/negativa-ai/BLAFS
We have tested the tool on many containers. For the top 20 pulled docker containers, the savings are up to 95% with all the containers working. We have done the same tests with slim-toolkit and it was only able to debloat 8 out of 20. We are looking at making this more automated for users.
Now, we only support docker, but we are working on podman and lXCs. We are also working on a version that guarantees generality (i.e., we give you 100% guarantees that your container will always work) with zero requirements to profile.
Please do try it and let us know what you think! All feedback is welcome!
