Forensicability Of Deep Neural Network Inference Pipelines

We propose methods to infer properties of the execution envi-ronment of machine learning pipelines by tracing characteris-tic numerical deviations in observable outputs. Results from aseries of proof-of-concept experiments obtained on local andcloud-hosted machines give raise to possible forensic applica-tions, such as the identification of the hardware platform usedto produce deep neural network predictions. Finally, we intro-duce boundary samples that amplify the numerical deviationsin order to distinguish machines by their predicted label only.