Hardened AI Ensemble Co-Training to Prevent Adversarial Transfer
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial intelligence models are vulnerable to adversarial attacks, particularly white box attacks that manipulate input data to deceive the model into incorrect classifications, and these defenses often focus on a single model, failing to address transferability of attacks to other models.
Innovation Solution
A hardened ensemble of AI models is created by co-training multiple models with a regularizer that modifies the loss surface to prevent adversarial attacks from transferring between models, ensuring that errors created for one model do not affect others, thereby enhancing robustness against adversarial samples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single AI model is used for classification, then the system is simple and easy to implement, but it becomes vulnerable to adversarial attacks and lacks robustness
Solution Approach 1:
The system segments the classification task into multiple independent AI models that each process the same input data. By dividing the single-model approach into an ensemble of diverse models with different architectures, training data, and hyperparameters, the system achieves robustness against adversarial attacks while maintaining manageable complexity through modular deployment
2Reliability
If multiple AI models are trained independently, then diversity is achieved, but adversarial attacks can transfer between models reducing overall effectiveness
Solution Approach 1:
The system implements feedback mechanisms where each model in the ensemble provides information about its predictions and confidence levels. This feedback loop allows the ensemble to detect when adversarial attacks are being applied and adjust individual model behaviors or weighting schemes to prevent attack transferability, making the training process more complex but significantly improving security
3Reliability
If adversarial training is applied to harden a model, then resistance to attacks improves, but the model may lose accuracy on benign inputs
Solution Approach 1:
The system applies local quality by treating each AI model in the ensemble differently - some models are specialized for adversarial robustness while others optimize for accuracy on benign inputs. By assigning different roles and training objectives to different models based on their local strengths, the ensemble maintains high classification accuracy on normal data while achieving strong adversarial resistance through the collective decision-making process
Data Source
AI summary
Mechanisms are provided to implement a hardened ensemble artificial intelligence (AI) model generator. The hardened ensemble AI model generator co-trains at least two AI models. The hardened ensemble AI model generator modifies, based on a comparison of the at least two AI models, a loss surface of one or more of the at least two AI models to prevent an adversarial attack on one AI model, in the at least two AI models, transferring to another AI model in the at least two AI models, to thereby generate one or more modified AI models. At least one of the one or more modified AI models then processes an input to generate an output result.