AI Model Poisoning Detection Using Causal Relationship Strength
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI models can be compromised by poisoned training data, leading to untrustworthy inferences and the need for computationally expensive re-training, which disrupts inference consumption and other services.
Innovation Solution
Analyze candidate training data sets by comparing causal relationships with historical data to identify potential poisoning, using a threshold to determine if the data is poisoned, thereby reducing the need for re-training and conserving computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If AI models are re-trained to remove poisoned training data, then model reliability is improved, but computational resources and time are significantly consumed
Solution Approach 1:
The patent applies preliminary action by detecting poisoned training data before re-training the AI model. The system analyzes training data sets to identify poisoned samples in advance, allowing the model to be re-trained only when necessary and using cleaned data, thus avoiding unnecessary re-training time while maintaining model reliability
Solution Approach 2:
The patent introduces an intermediary detection mechanism between data collection and model training. This intermediary system analyzes training data sets to identify poisoned samples, acting as a mediator that filters harmful data before it reaches the training process, thereby preventing reliability issues without requiring full re-training
2Reliability
If AI models are re-trained to remove poisoned training data, then model reliability is improved, but computational resources are consumed
Solution Approach 1:
The system performs preliminary detection of poisoned training data before initiating re-training. By identifying and removing poisoned samples in advance through data set analysis, the system ensures that re-training is only performed when necessary and with cleaned data, thereby improving model reliability while minimizing computational resource consumption
Solution Approach 2:
An intermediary detection system is introduced that analyzes training data sets to identify poisoned samples. This intermediary mechanism filters harmful data before training, acting as a protective layer that maintains model reliability without requiring computationally expensive full re-training cycles
3Reliability
If re-training is performed frequently to address poisoned data, then model reliability is improved, but inference consumption and services are disrupted
Solution Approach 1:
The system performs preliminary detection of poisoned training data before re-training is initiated. By identifying poisoned samples in advance through analysis of training data sets, the system can clean the data and perform targeted re-training only when necessary, thereby maintaining model reliability while minimizing disruption to inference consumption and services
Solution Approach 2:
An intermediary detection mechanism is introduced between data collection and model training. This intermediary system analyzes training data sets to identify poisoned samples, serving as a filter that prevents harmful data from compromising the model. This approach maintains reliability while avoiding frequent, disruptive re-training cycles
Data Source
AI summary
Methods and systems for managing artificial intelligence (AI) models are disclosed. To manage AI models, an instance of an AI model may not be re-trained using training data determined to be potentially poisoned. By doing so, malicious attacks intending to influence the AI model using poisoned training data may be prevented. To do so, a first level of strength of a first causal relationship present in historical training data may be compared to a second level of strength of a second causal relationship present in a candidate training data set. The first level of strength and the second level of strength may be expected to be similar within a threshold. If a difference between the first level of strength and the second level of strength is not within the threshold, the candidate training data may be treated as including poisoned training data.