Selective Data Poisoning for Generative Model Classification Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for training generators in generative adversarial networks (GANs) do not effectively control the accuracy and specificity of generated data, leading to generators that may not be suitable for intended tasks due to misclassification in certain labeled data sets.
Innovation Solution
A method is introduced to create a selectively poisoned generator by altering specific data in the training set, training a discriminator, and then retraining the generator to achieve desired levels of misclassification for unauthorized labels while maintaining accuracy for authorized labels, using a processor to input and alter data, train the discriminator, and test the generator's performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the generator is trained to improve accuracy for authorized labels, then classification accuracy for authorized labels is improved, but the generator may inadvertently improve accuracy for unauthorized labels as well
Solution Approach 1:
The patent applies local quality by selectively poisoning only specific portions of the training data corresponding to unauthorized labels, while leaving authorized label data unaltered. This creates localized degradation in specific classification pathways while preserving accuracy in authorized pathways, enabling the generator to maintain high accuracy for authorized labels while deliberately reducing accuracy for unauthorized labels.
Solution Approach 2:
The training data is segmented into authorized and unauthorized portions, with different treatment applied to each segment. The poisoning process is applied selectively to the unauthorized segment, creating distinct classification performance characteristics for different label types. This segmentation enables independent control over accuracy for different label categories.
2Object-generated harmful factors
If data is altered in the training set to poison the generator, then misclassification for unauthorized labels is improved, but accuracy for authorized labels may deteriorate
Solution Approach 1:
The poisoning alteration is applied locally only to data portions corresponding to unauthorized labels, while authorized label data remains completely unaltered. This selective application ensures that harmful misclassification effects are concentrated only where intended (unauthorized labels) while authorized label classification reliability is preserved at full strength.
Solution Approach 2:
The poisoning process is performed as a preliminary step before generator training, where specific training data is pre-altered to encode the desired misclassification behavior. This preliminary action embeds the selective poisoning characteristics into the training data itself, ensuring that the generator learns the desired discriminatory accuracy patterns from the outset without requiring post-training adjustments that might affect authorized labels.
3Measurement precision
If the generator is made highly accurate for all labels, then overall classification accuracy is improved, but the ability to selectively control classification for specific tasks is lost
Solution Approach 1:
The patent changes the parameter distribution of the training data by introducing selective poisoning alterations that modify the accuracy parameter for specific label categories. By adjusting the degree and nature of data alterations for unauthorized labels while maintaining original data for authorized labels, the system creates a generator with differentiated accuracy parameters across different label types, enabling controlled use and pricing strategies.
Solution Approach 2:
The patent converts the harmful effect of data poisoning (which typically degrades overall model performance) into a beneficial feature by selectively applying poisoning only to unauthorized label data. This transforms what would normally be a detrimental accuracy reduction into a useful mechanism for task-specific control, allowing the generator to be optimized for specific authorized tasks while deliberately performing poorly on unauthorized tasks.
Data Source
AI summary
A generator for generating artificial data, and training for the same. Data corresponding to a first label is altered within a reference labeled data set. A discriminator is trained based on the reference labeled data set to create a selectively poisoned discriminator. A generator is trained based on the selectively poisoned discriminator to create a selectively poisoned generator. The selectively poisoned generator is tested for the first label and tested for the second label to determine whether the generator is sufficiently poisoned for the first label and sufficiently accurate for the second label. If it is not, the generator is retrained based on the data set including the further altered data. The generator includes a first ANN to input first information and output a set of artificial data that is classifiable using a first label and not classifiable using a second label of the set of labeled data.