Secure industrial federated learning: Label encryption for model protection

Amid the widespread adoption of deep learning techniques in the Industrial Internet of Things, unauthorized extraction of trained deep learning models has emerged as a significant threat to model intellectual property protection. Existing intellectual property protection methods, such as neural network watermarking and fingerprinting, primarily focus on passive tracing, lacking proactive prevention capabilities. In this paper, we propose an encryption-based intellectual property protection framework specifically designed for federated learning scenarios in industrial settings. The primary artificial intelligence-related contribution of this framework is the design of an efficient label encryption scheme, which selectively encrypts only label information rather than entire datasets or model parameters, significantly reducing computational and communication overhead while preserving model accuracy. The primary engineering application involves integrating encryption mechanisms into the federated learning training and deployment processes to ensure proactive access control and robust passive traceability. The proposed framework employs a hierarchical access control protocol leveraging client-specific encryption keys, providing active prevention against unauthorized model use and facilitating forensic evidence collection. Additionally, the encryption mechanism protects client data privacy and clearly establishes model ownership. Through comprehensive experiments and analyses, we demonstrate that the proposed encryption-based framework effectively safeguards model intellectual property, preserves model performance, and achieves robustness suitable for resource-constrained industrial environments.