Towards Privacy-Preserving Machine Learning for Energy Prediction in Industrial Robotics: Modeling, Evaluation and Integration

Adam Skuta, Philipp Steurer, Sebastian Hegenbart, Ralph Hoch, Thomas Loruenser

This paper explores the feasibility and implications of developing a privacy-preserving, data-driven cloud service for predicting the energy consumption of industrial robots. Using machine learning, we evaluated three neural network architectures—dense, LSTM, and convolutional–LSTM hybrids—to model energy usage based on robot trajectory data. Our results show that models incorporating manually engineered features (angles, velocities, and accelerations) significantly improve prediction accuracy. To ensure secure collaboration in industrial environments where data confidentiality is critical, we integrate privacy-preserving machine learning (ppML) techniques based on secure multi-party computation (SMPC). This allows energy inference to be performed without exposing proprietary model weights or confidential input trajectories. We analyze the performance impact of SMPC on different network types and evaluate two optimization strategies, using public model weights through permutation and evaluating activation functions in plaintext, to reduce inference overhead. The results highlight that network architecture plays a larger role in encrypted inference efficiency than feature dimensionality, with dense networks being the most SMPC-efficient. In addition to model development, we identify and discuss specific stages in the MLOps workflow—particularly model serving and monitoring—that require adaptation to support ppML. These insights are useful for integrating ppML into modern machine learning pipelines.