On Improving Deep Active Learning with Formal Verification

TL;DR

This paper explores how integrating adversarial inputs generated through formal verification into deep active learning enhances model robustness and data efficiency, outperforming traditional gradient-based adversarial methods.

Contribution

It introduces a novel approach of using formally verified adversarial examples to improve deep active learning, demonstrating significant performance gains over existing methods.

Findings

Formally verified adversarial examples outperform gradient-based attacks in improving DAL.

Augmenting training with these adversarial inputs enhances model generalization.

The proposed method yields significant improvements across multiple benchmarks.

Abstract

Deep Active Learning (DAL) aims to reduce labeling costs in neural-network training by prioritizing the most informative unlabeled samples for annotation. Beyond selecting which samples to label, several DAL approaches further enhance data efficiency by augmenting the training set with synthetic inputs that do not require additional manual labeling. In this work, we investigate how augmenting the training data with adversarial inputs that violate robustness constraints can improve DAL performance. We show that adversarial examples generated via formal verification contribute substantially more than those produced by standard, gradient-based attacks. We apply this extension to multiple modern DAL techniques, as well as to a new technique that we propose, and show that it yields significant improvements in model generalization across standard benchmarks.