Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
102 tokens/sec
GPT-4o
59 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
50 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Robust Pre-Training by Adversarial Contrastive Learning (2010.13337v1)

Published 26 Oct 2020 in cs.CV

Abstract: Recent work has shown that, when integrated with adversarial training, self-supervised pre-training can lead to state-of-the-art robustness In this work, we improve robustness-aware self-supervised pre-training by learning representations that are consistent under both data augmentations and adversarial perturbations. Our approach leverages a recent contrastive learning framework, which learns representations by maximizing feature consistency under differently augmented views. This fits particularly well with the goal of adversarial robustness, as one cause of adversarial fragility is the lack of feature invariance, i.e., small input perturbations can result in undesirable large changes in features or even predicted labels. We explore various options to formulate the contrastive task, and demonstrate that by injecting adversarial perturbations, contrastive pre-training can lead to models that are both label-efficient and robust. We empirically evaluate the proposed Adversarial Contrastive Learning (ACL) and show it can consistently outperform existing methods. For example on the CIFAR-10 dataset, ACL outperforms the previous state-of-the-art unsupervised robust pre-training approach by 2.99% on robust accuracy and 2.14% on standard accuracy. We further demonstrate that ACL pre-training can improve semi-supervised adversarial training, even when only a few labeled examples are available. Our codes and pre-trained models have been released at: https://github.com/VITA-Group/Adversarial-Contrastive-Learning.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (4)
  1. Ziyu Jiang (16 papers)
  2. Tianlong Chen (202 papers)
  3. Ting Chen (148 papers)
  4. Zhangyang Wang (375 papers)
Citations (215)

Summary

An Overview of "Robust Pre-Training by Adversarial Contrastive Learning"

In the current landscape of machine learning research, robustness and label efficiency are crucial objectives for deep neural network training. The paper entitled "Robust Pre-Training by Adversarial Contrastive Learning" by Jiang et al. presents a novel approach that integrates adversarial perturbations with contrastive learning to enhance model robustness during pre-training. This paper aims to achieve models that are both robust in adversarial settings and efficient in terms of label usage.

Methodological Advancements

The authors build upon the recent progress in contrastive learning frameworks, particularly the SimCLR model, to propose an approach known as Adversarial Contrastive Learning (ACL). The core idea of ACL is to leverage adversarial perturbations within a contrastive learning setup. This methodology seeks to maximize feature consistency under adversarial circumstances, aiming to enhance both standard and robust accuracies.

Different strategies for integrating adversarial components into contrastive learning are explored, resulting in three distinct variants:

  1. Adversarial-to-Adversarial (A2A): This variant involves applying adversarial attacks to both branches of contrastive pairs after standard augmentations. However, it demonstrated only minor robustness gains while diminishing standard accuracy due to its aggressive consistency enforcement.
  2. Adversarial-to-Standard (A2S): Here, only one branch receives adversarial perturbations, while the second remains as a standard augmentation. This approach achieves better robustness while maintaining standard accuracy.
  3. Dual Stream (DS): Combining both S2S and A2A strategies, this dual approach optimizes over both standard and adversarial data. It showed the best results, significantly improving both the adversarial and standard testing accuracies.

Empirical Evidence

The paper provides extensive empirical evaluations on popular datasets like CIFAR-10 and CIFAR-100. Notably, the ACL approach surpasses previous state-of-the-art methods in both robust and standard accuracies. For example, on CIFAR-10, ACL outperformed the leading unsupervised robust pre-training approach with enhancements of 2.99% in robust accuracy and 2.14% in standard accuracy.

Moreover, the ACL pre-training was found to significantly benefit semi-supervised learning scenarios, particularly in low-label regimes. When only 1% of the CIFAR-10 labels were available, ACL maintained a high robustness, showing the method's potential in data-efficient learning conditions.

Implications and Future Directions

By proposing ACL, the authors bridge a crucial gap in leveraging contrastive learning for robust model development. The introduction of adversarial components into contrastive frameworks not only provides resilience against adversarial attacks but also enhances general standard learning objectives.

This work advances the field by offering a pre-training paradigm that could be pivotal for various downstream tasks, not limited to adversarial circumstances. Future research might investigate the scalability of this approach to larger models and datasets, explore different types of adversarial perturbations, or integrate this framework into other unsupervised learning paradigms.

In conclusion, "Robust Pre-Training by Adversarial Contrastive Learning" presents a compelling case for the integration of adversarial learning and contrastive representations, setting a new benchmark in the design of robust and label-efficient neural models.