Causal Analysis for Detecting Poisoned AI Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
AI models can become untrustworthy if poisoned training data is introduced, leading to computationally expensive re-training processes that divert resources from inference generation.
Innovation Solution
A system that analyzes candidate training data sets by comparing them to historical data sets to detect poisoned data, allowing for the determination of whether to update an AI model instance without re-training if the data is deemed poisoned.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If AI models are re-trained to remove the influence of poisoned training data, then the trustworthiness of AI model outputs is improved, but computational resources and time are significantly consumed
Solution Approach 1:
The system performs preliminary detection of poisoned training data by comparing causal relationships in candidate training data against historical training data before the AI model is actually trained. This advance detection prevents the model from being exposed to poisoned data in the first place, eliminating the need for costly re-training while maintaining model trustworthiness.
Solution Approach 2:
The system applies preliminary anti-action by identifying and blocking poisoned training data before it can influence the AI model. By detecting causal relationship discrepancies in advance, the system neutralizes the harmful effect of poisoned data, preventing the need for corrective re-training actions.
2Reliability
If AI models are re-trained to remove the influence of poisoned training data, then the trustworthiness of AI model outputs is improved, but disruptions to inference consumption increase
Solution Approach 1:
The system performs preliminary detection of poisoned training data before the AI model training process begins. By identifying causal relationship discrepancies in advance, the system prevents poisoned data from entering the training pipeline, thereby avoiding re-training disruptions and maintaining continuous inference consumption without interruption.
3Measurement precision
If candidate training data sets are analyzed by comparing causal relationships to historical data sets, then detection accuracy of poisoned data is improved, but analysis complexity increases
Solution Approach 1:
The system introduces causal relationships as an intermediary layer between raw training data and poisoned data detection. By analyzing causal relationships rather than raw data directly, the system achieves high detection accuracy while simplifying the analysis process, as causal relationships provide a structured framework for comparison against historical data.
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 influence the AI model in a using poisoned training data may be prevented. To do so, a first causal relationship present in historical training data may be compared to a second causal analysis present in a candidate training data set. The first causal relationship and the second causal relationship may be expected to be similar within a threshold. If a difference between the first causal relationship and the second causal relationship is not within the threshold, the candidate training data may be treated as including poisoned training data.