Privacy, Trust & Intelligent Systems in Industry 6.0

Industrial sites want to learn from each other's data without disclosing it and to auto-mate decisions only where errors stay within stated limits. Private federated detectors report accuracy and privacy budgets but give no certified bound on the decisions that follow. We propose private aggregate risk control (PARC): sites train a model with dif-ferentially private federated averaging, add Skellam noise to histograms of calibration scores and combine them by secure aggregation and the server certifies automatic-pass and automatic-alarm thresholds with a binomial-Skellam test. We prove finite-sample validity under the privacy noise and derive the minimum calibration evidence that certification needs. On eight industrial and security tasks (198,028 observations; 95 confirmatory runs) PARC exceeded a 10% miss-rate target in 1 of 325 runs against 39 for a private plug-in rule. Private calibration was nearly free at ε ≥ 4, but a 10% miss-rate certificate needed at least 31 calibration faults without privacy and 118 at ε = 1. DP training cut rare-fault PR-AUC on AI4I from 0.581 to 0.264 at ε = 4 and drift broke interval coverage on a gas turbine. Operators should count failure evidence be-fore choosing privacy budgets and treat escalation as the channel for human oversight.