Sanitizing Machine Learning Models via Path Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models are vulnerable to adversarial attacks due to poisoned training data, making it difficult to detect and mitigate adversarial biases without significant additional data or complete retraining, especially when pre-trained models with unknown behavior are used.
Innovation Solution
A method is provided to sanitize machine learning models by training a path between a potentially poisoned model and a clean model using limited clean training data, generating a sanitized model that is less prone to misclassification, and reducing adversarial risk while maintaining accuracy, by determining a set of weights that minimize loss along a continuous piece-wise smooth parametric curve.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If pre-trained models are used to reduce training data requirements, then productivity is improved, but reliability deteriorates due to potential poisoning and adversarial biases
Solution Approach 1:
The patent introduces a path training process as an intermediary mechanism between the pre-trained model and final deployment. This path acts as a mediator that gradually transitions the model from its pre-trained state to a sanitized state, eliminating adversarial biases while preserving useful learned features. The path training serves as a buffer that reconciles the conflict between using pre-trained models for efficiency and ensuring their reliability.
2Reliability
If additional training data is collected to retrain models, then reliability is improved, but loss of time and productivity deteriorate
Solution Approach 1:
The patent applies partial action by performing path training on only a subset of training data rather than complete retraining with full datasets. The path training process selectively adjusts model parameters along a continuous path using limited clean data, eliminating adversarial biases without requiring extensive additional data collection and processing time.
3Productivity
If pre-trained models are used without sanitization, then productivity is improved, but object-affected harmful factors increase due to adversarial attacks
Solution Approach 1:
The patent implements preliminary action by performing path training and sanitization before deploying the model to production environments. The path training process proactively eliminates adversarial biases and poisonings in advance, ensuring that models are hardened against future attacks before they encounter real-world adversarial inputs, thus maintaining both productivity and security.
4Reliability
If path training is performed to sanitize models, then reliability is improved, but use of energy and computational resources increases
Solution Approach 1:
The patent applies parameter changes by modifying model parameters along a continuous path during path training rather than performing exhaustive retraining. The path training process efficiently adjusts parameters through gradient-based optimization along a predefined trajectory, achieving sanitization with reduced computational overhead compared to complete retraining while maintaining improved reliability.
Data Source
AI summary
Techniques for sanitization of machine learning (ML) models are provided. A first ML model is received, along with clean training data. A path is trained between the first ML model and a second ML model using the clean training data. A sanitized ML model is generated based on at least one point on the trained path. One or more ML functionalities are then facilitated using the sanitized ML model.