AugLy: Data Augmentations for Robustness

TL;DR

AugLy introduces a multimodal data augmentation library focused on robustness to real-world online perturbations across audio, image, text, and video. It offers a unified API, metadata support, and multimodal composition to simulate realistic perturbations and evaluate robustness beyond traditional training-time augmentations. The paper benchmarks AugLy against modality-specific libraries and demonstrates its utility by conducting robustness evaluations on ImageNet models, revealing vulnerability patterns and the impact of different augmentation strategies. This work enables systematic robustness assessment and adversarial-like data generation with an accessible, open-source tool.

Abstract

We introduce AugLy, a data augmentation library with a focus on adversarial robustness. AugLy provides a wide array of augmentations for multiple modalities (audio, image, text, & video). These augmentations were inspired by those that real users perform on social media platforms, some of which were not already supported by existing data augmentation libraries. AugLy can be used for any purpose where data augmentations are useful, but it is particularly well-suited for evaluating robustness and systematically generating adversarial attacks. In this paper we present how AugLy works, benchmark it compared against existing libraries, and use it to evaluate the robustness of various state-of-the-art models to showcase AugLy's utility. The AugLy repository can be found at https://github.com/facebookresearch/AugLy.

Figures (20)

• Figure 1: Examples of a few AugLy image augmentations
• Figure 2: Examples of some AugLy text augmentations
• Figure 3: Calling an image augmentation
• Figure 4: Composing audio & video augmentations
• Figure 5: The audio augmentation libraries we chose to compare and their corresponding number of augmentations at the time of writing.