
Jonas Möller
Doctoral Researcher
Jonas Möller, Lukas Pirch, Felix Weissberg, Sebastian Baunsgaard, Thorsten Eisenhofer, Konrad Rieck
Adversarial Inputs for Linear Algebra Backends
July
13,
2025
https://www.mlsec.org/docs/2025-icml.pdf

Machine Learning| Jul 17, 2025
Even the smallest number can make a big difference
Minor deviations in backend libraries like CUDA or MKL can cause identical AI models to produce different outputs. At ICML 2025, BIFOLD researcher Konrad Rieck showed how such subtle imprecisions can be exploited—posing a significant risk to AI system security.