Golden Data Validation for ML Model Resiliency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning workloads face challenges in maintaining the reliability and integrity of models and weights due to corruption or errors, especially as models and weights increase in size, leading to performance issues and the need for improved resiliency mechanisms.
Innovation Solution
The use of 'golden data' – a smaller data structure comprising inputs and expected outputs – is employed to validate the stability of machine learning models and weights, stored in secure memory and decrypted for validation, allowing for periodic or request-based validation to ensure model integrity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the size of machine learning models and weights increases to improve model capability, then model performance is improved, but the complexity of validation and storage requirements increase
Solution Approach 1:
The patent segments the validation process into two distinct components: (1) validation of model weights using cryptographic signatures, and (2) validation of model outputs using golden test cases. This segmentation allows each validation aspect to be handled independently with appropriate methods, reducing overall validation complexity despite increasing model size.
Solution Approach 2:
The patent implements preliminary validation actions by verifying cryptographic signatures of model weights before deployment and by pre-establishing golden test cases with expected outputs. These preliminary checks are performed before the model processes actual workloads, enabling early detection of corruption or errors without requiring complex runtime validation of large models.
2Reliability
If encryption is applied to protect model and weight integrity, then security is improved, but processing performance deteriorates
Solution Approach 1:
The patent segments protection mechanisms into different layers: cryptographic signatures protect weight integrity during storage and loading, while golden test cases validate output correctness during inference. This segmentation allows encryption to be applied only where necessary (weight storage) rather than to entire model processing, minimizing performance impact while maintaining reliability.
Solution Approach 2:
The patent creates cryptographic copies (signatures) of model weights that can be verified without decrypting or processing the actual weight data. This copying approach allows integrity verification to be performed independently from the main model processing pipeline, preventing encryption/decryption operations from bottlenecking overall processing performance.
3Reliability
If validation is performed frequently to ensure model stability, then reliability is improved, but computational overhead increases
Solution Approach 1:
The patent implements periodic validation using golden test cases at scheduled intervals or at specific trigger points (e.g., after weight updates, before deployment, or at configured frequencies). This periodic approach ensures model stability is maintained through regular checks while avoiding continuous validation that would create excessive computational overhead and energy consumption.
Solution Approach 2:
The patent extracts a small subset of critical validation cases (golden test cases) from the full model validation process. These extracted test cases focus on key functionality and can be executed quickly with minimal computational resources, enabling frequent validation checks without proportionally increasing computational overhead.
Data Source
AI summary
In exemplary aspects, a golden data structure can be used to validate the stability of machine learning (ML) models and weights. The golden data structure includes golden input data and corresponding golden output data. The golden output data represents the known correct results that should be output by a ML model when it is run with the golden input data as inputs. The golden data structure can be stored in a secure memory and retrieved for validation separately or together with the deployment of the ML model for a requested ML operation. If the golden data structure is used to validate the model and/or weights concurrently with the performance of the requested operation, the golden input data is combined with the input data for the requested operation and run through the model. Relevant outputs are compared with the golden output data to validate the stability of the model and weights.