AI Model Poisoning Remediation via Partial Re-Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
AI models can become untrustworthy due to poisoned training data, leading to inaccurate inferences and requiring costly re-training, which interrupts inference generation and consumes significant computing resources, energy, and network bandwidth.
Innovation Solution
The system manages AI models by periodically taking snapshots during training, allowing for the restoration of a partially trained untainted model using a subset of the training dataset, and determines whether to remediate poisoned inferences based on their potential impact, reducing the need for full re-training and resource expenditure.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If full re-training of the AI model is performed to remove poisoned data influence, then model reliability is improved, but computing resources, energy consumption, and network bandwidth are significantly consumed
Solution Approach 1:
The patent segments the training dataset into portions and identifies poisoned portions separately. Instead of re-training with the entire dataset, only the poisoned portions are removed and re-training proceeds with the clean portions, significantly reducing computational resources while maintaining model reliability.
Solution Approach 2:
The patent extracts and removes only the poisoned training data portions from the complete training dataset. This extraction approach allows the system to eliminate the harmful influence of poisoned data without processing or re-training on the entire dataset, thereby conserving computing resources and energy.
2Reliability
If full re-training of the AI model is performed to remove poisoned data influence, then model reliability is improved, but inference generation is interrupted
Solution Approach 1:
The patent applies partial action by performing re-training only on the portions of the training dataset that are affected by poisoning, rather than re-training with the complete dataset. This partial re-training approach maintains model reliability while minimizing interruption to inference generation services.
Solution Approach 2:
By segmenting the training process and identifying only the poisoned data portions that require re-training, the system can perform targeted updates without halting inference generation for extended periods, thus maintaining productivity while improving reliability.
3Measurement precision
If poisoned inferences are remediated by generating replacement inferences, then inference accuracy is improved, but network bandwidth and computing resources are consumed
Solution Approach 1:
The patent extracts and identifies only the poisoned inferences that require remediation, rather than regenerating all inferences. By targeting only the affected portion, the system improves inference accuracy while minimizing network bandwidth consumption associated with transmitting replacement inferences.
Data Source
AI summary
Methods and systems for managing artificial intelligence (AI) models are disclosed. Training data used to train an instance of an AI model may be poisoned, thereby generating a poisoned AI model. Inferences generated using the poisoned AI model may be poisoned inferences and poisoned inferences may negatively impact operation of an inference consumer. The potential impact of the poisoned inferences on the inference consumer may be evaluated to determine whether the remediate the poisoned inferences. If the potential impact of the poisoned inferences is considered unacceptable, the poisoned inferences may be remediated by notifying the inference consumer of the poisoned inferences. The poisoned inferences may also be remediated by replacing the poisoned inferences with updated inferences generated using an updated instance of the AI model not trained using the poisoned training data.