Mitigating adversarial effects in machine learning systems

Abstract

Techniques for cleaning a machine learning (ML) model are provided. A first ML model and clean training data are received. A path between the first ML model and the second ML model is trained using the clean training data. A scrubbed ML model is generated based on at least one point on the trained path. One or more ML functions are then facilitated using the purged ML model.

Description

technical field [0001] The present disclosure relates to machine learning, and more specifically, to mitigating adversarial effects in machine learning systems. Background technique [0002] Various machine learning models and techniques have been used in various applications and continue to show great potential. However, many models require large amounts of training data in order to return satisfactory results. Obtaining such data is often difficult, expensive and time-consuming. To reduce these burdens, many rely on pretrained models or prelabeled training data. However, these models and data are relatively unknown and could be contaminated or poisoned. For example, training data may be corrupted by specified trigger patterns associated with target labels. During training, these models incorporate this adversarial behavior such that when specified triggers are present, these models can be tricked or forced to misclassify input data into the target category. Other atta...

AI Technical Summary

Problems solved by technology

Often, these adversarial biases are difficult or impossible to detect

Images