Systems and methods for adaptively improving the performance of locked machine learning programs
Inventors
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-12333393-B2
Publication Date
2025-06-17
Expiration Date
Abstract
Techniques for adaptively improving the performance of a locked machine learning program have been disclosed. In one particular embodiment, the techniques may be realized as a method for enabling a first party to provide a trained machine learning model to a second party, the method comprising receiving the trained machine learning model from the first party, the trained machine learning model being associated with one or more policies defining permissible operations; and constraining the second party to operate the trained machine learning model in a manner that is consistent with said one or more policies.
Core Innovation
The invention enables a first party to provide a trained machine learning model to a second party while preserving control via one or more policies defining permissible operations. The trained machine learning model is received from the first party and is associated with the one or more policies, and the second party is constrained to operate the trained machine learning model consistently with the one or more policies.
A program operating in a secure enclave serving as a nth-stage of a secure pipeline collects a dataset and transmits the dataset to a (n−1) stage of the secure pipeline. The trained machine learning model is re-trained using the dataset, and the model is locked with respect to a first demographic prior to re-training.
After re-training, the trained machine learning model is locked with respect to a heuristically generated different demographic associated with the dataset. In the described environment, encrypted operational log output supports policy compliance, and authenticated clients can add remarks that produce a retraining dataset returned to the provider for decryption within a secure enclave.
The documented approach includes de-identifying PHI and using heuristic set manipulation to create demographic subsets distinct from the original locked training set, followed by re-training to broaden or maintain demographic performance. Secure pipeline stages and verifiable audit logs including digital certificates/attestation records support regulator verification and digital trust through keys integrity.
Claims Coverage
The independent claims are directed to method, system, and non-transitory computer readable medium implementations of the same core workflow, comprising secure policy-constrained model usage, secure pipeline/enclave dataset handling with stage-wise transmission, and demographic-based locking before and after re-training. Across the independent claims, three primary inventive features are recited.
Policy-associated trained model constrained by permissible operations
The trained machine learning model is received from the first party, associated with one or more policies defining permissible operations, and the second party is constrained to operate the trained machine learning model in a manner consistent with the one or more policies.
Secure enclave nth-stage collection and (n−1) stage dataset transmission
A program operating in a secure enclave serving as a nth-stage of a secure pipeline collects a dataset and transmits the dataset to a (n−1) stage of the secure pipeline.
Demographic-locked re-training using the dataset with heuristically generated different demographic
The trained machine learning model is re-trained using the dataset, where the trained machine learning model is locked with respect to a first demographic prior to re-training and is locked with respect to a heuristically generated different demographic associated with the dataset after re-training.
Each independent claim covers the combination of policy-defined permissible operations that constrain second-party use, secure enclave/secure pipeline dataset handling with stage transmission, and re-training in which the model is locked to a first demographic before re-training and locked to a heuristically generated different demographic associated with the dataset after re-training.
Stated Advantages
Enables locked machine learning program improvement across parties while preserving privacy/IP.
Supports policy compliance and regulator verification via verifiable audit logs, digital certificates, and attestation records.
Uses demographic subset creation to broaden or maintain demographic performance.
Outputs encrypted operational logs to support compliance.
Documented Applications
Adaptive improvement of a locked ML program across parties in an SaMD ecosystem (including parties such as diagnostic center).
Provider-side retraining using an operational log–derived retraining dataset with demographic performance maintenance/broadening.
Interested in licensing this patent?