How Can Industrial Robots Predict Their Energy Use Without Revealing Secrets?

Imagine a robot that knows how much energy it will use before it even moves. Researchers trained three neural network types – dense, LSTM, and convolutional‑LSTM hybrids – using robot trajectory data. They added manual features like angles, speeds, and accelerations, making predictions more accurate. But companies don’t want to share sensitive data. To solve this, they introduced privacy‑preserving machine learning techniques based on secure multi‑party computation (SMPC). With SMPC, the robot can ask the cloud to estimate consumption without revealing its movements or model weights. Experiments show dense networks are the fastest under encryption, while data complexity matters less. The authors also tested two tricks: using permuted public weights and computing activation functions in clear, cutting inference time. Finally, they highlighted that serving and monitoring stages of the MLOps pipeline must be adapted to handle encryption. The result is a cloud service that can predict robot energy use while keeping industrial secrets safe.

The study relies on Secure Multi‑Party Computation (SMPC), a family of cryptographic protocols that enable multiple parties to jointly compute a function without revealing their private inputs. In practice, robot trajectory data (positions, speeds, accelerations) are split into secret shares and distributed across several servers. Each server performs arithmetic on its shares; only after combining the results does the energy consumption prediction emerge. This eliminates the need to transmit raw data or model weights in clear. Regarding architectures, dense networks consist of fully connected layers, reducing the number of non‑linear encrypted operations, making them more efficient under SMPC. LSTMs, however, involve many temporal steps and sigmoid/tanh gates, which increase computational cost when encrypted. Convolutional‑LSTM hybrids combine convolutions (which can be executed in clear with optimizations) and LSTMs, but the recurrent part remains the bottleneck. To mitigate overhead, the authors proposed two strategies: (1) permuting public weights so the model does not expose proprietary information, and (2) evaluating activation functions in plaintext, leveraging their non‑sensitive nature. These techniques drastically cut the number of cryptographic operations while preserving comparable accuracy. Moreover, integrating ppML into the MLOps workflow demands changes to model serving: containers must support SMPC protocols, and monitoring must capture cryptographic latency metrics alongside accuracy. Secret management, key rotation, and share integrity verification become essential parts of the model lifecycle. In summary, the approach shows that predictive performance, intellectual property protection, and cloud scalability can coexist, paving the way for more secure AI services in Industry 4.0.