Variant Inconsistency Attack for Adversarial Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing adversarial attack methods require labeled data, which is hard to procure in real-world applications, limiting the generation of adversarial examples for adversarial training of machine-learning models.
Innovation Solution
The Variant Inconsistency Attack (VIA) method generates adversarial examples by modifying input data in a utility-preserving manner without changing the expected label, allowing for the creation of adversarial pairs without labeled data, using techniques like adding noise or performing insert, delete, or replace operations on text data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing adversarial attack methods are used, then adversarial examples can be generated, but labeled data is required which is hard to procure in real-world applications
Solution Approach 1:
The system generates adversarial examples autonomously by detecting prediction inconsistencies across multiple model runs without requiring external labeled data. The model serves itself by using its own prediction variations to identify and generate adversarial examples, eliminating the need for human-labeled datasets.
Solution Approach 2:
The method changes the parameter of using unlabeled data instead of labeled data for generating adversarial examples. By modifying the input data requirements from labeled to unlabeled, the system makes adversarial attack generation feasible in real-world scenarios where labeled data is scarce or expensive to obtain.
2Productivity
If utility-preserving modifications are applied to generate adversarial pairs, then unlimited adversarial examples can be produced, but the process requires identifying prediction inconsistencies
Solution Approach 1:
The system performs preliminary actions by running multiple predictions on modified examples before finalizing adversarial example generation. By pre-comparing predictions across multiple runs and identifying inconsistencies early, the system efficiently filters out non-adversarial examples and focuses on generating genuine adversarial pairs.
Solution Approach 2:
The method incorporates feedback loops where prediction results from multiple runs are compared and used to guide subsequent adversarial example generation. The inconsistency detection mechanism provides feedback on which modifications successfully create adversarial examples, allowing the system to iteratively improve and scale its generation process.
Data Source
AI summary
The present disclosure relates to techniques for using variant inconsistency attack (VIA) as a simple and effective adversarial attack method to create useful adversarial examples for adversarial training of machine-learning models. In one particular aspect, a method is provided that includes obtaining a set of input examples for attacking a machine-learning model (the set of examples do not have corresponding labels), modifying an example from the set of examples in a utility preserving manner to generate a pair of modified examples, attacking the machine-learning model with the pair of modified examples in order generate a pair of predictions for the pair of modified examples, comparing the pair of predictions to determine whether the pair of predictions are the same or different, and in response to the pair of predictions being different, adding the pair of modified examples to a set of adversarial examples.