Robustness-preserving Lifelong Learning via Dataset Condensation

Jinghan Jia, Yihua Zhang, Dogyoon Song, Sijia Liu, Alfred O. Hero Iii

Abstract

Lifelong learning (LL) models often suffer from catastrophic forgetting and adversarial vulnerability. While memory replay is a common solution for maintaining accuracy, preserving robustness is often overlooked or requires excessive unlabeled data. This paper proposes DERPLL (Data-Efficient Robustness-Preserving Lifelong Learning), a framework that employs bi-level optimization to select an optimal coreset for replay. By integrating a robustness-aware training objective, DERPLL maintains both high standard accuracy and adversarial robustness across sequential tasks without relying on external data.

Introduction

Machine learning models deployed in dynamic environments face two primary challenges: catastrophic forgetting and adversarial vulnerability. In lifelong learning scenarios, models tend to lose performance on previous tasks as they adapt to new ones. Furthermore, they remain susceptible to adversarial attacks—small, intentional perturbations that cause incorrect predictions. While memory replay (storing a subset of old data) helps mitigate forgetting, existing methods typically prioritize standard accuracy over security. DERPLL addresses this by identifying a highly representative "coreset" of data to preserve both accuracy and robustness efficiently.

Methodology

The DERPLL framework consists of two core components designed to optimize the memory replay process:

1. Bi-level Optimization for Coreset Selection

Rather than using random sampling, the authors utilize Bi-level Optimization (BLO) to select a subset of data (a coreset). The objective is to find a subset such that a model trained exclusively on it performs similarly to one trained on the full dataset. To overcome the discrete nature of sample selection, the framework employs a "Straight-Through" gradient estimator and gradient unrolling, making the optimization feasible for deep architectures like ResNet-18.

2. Robustness-Aware Training Objective

During the learning of new tasks, DERPLL employs a composite loss function:

  • Adversarial Training (AT): Ensures robustness on the current task by training on adversarial examples.
  • Robust Learning without Forgetting (LwF): Uses the selected coreset to ensure that the model's predictions on both clean and perturbed data do not deviate significantly from the previous model's state.

Experimental Results

The researchers evaluated DERPLL on the CIFAR-10 dataset in a class-incremental setting, splitting the data into five distinct tasks. The results demonstrate that DERPLL consistently outperforms baseline methods like random selection and Influence Score (IS) selection.

Table 1: Performance comparison (Standard Accuracy and Robust Accuracy) on CIFAR-10 across sequential tasks.

As shown in the table above, DERPLL achieves superior preservation of both Standard Accuracy (SA) and Robust Accuracy (RA) as the number of tasks increases. In a 5-task sequence, it outperforms baselines by approximately 4% in SA and over 6% in RA on older tasks.

Data Efficiency and Scaling

A key advantage of DERPLL is its performance in data-constrained environments. The effectiveness of the bi-level optimization becomes more pronounced as the size of the memory bank decreases.

Figure 1: Robust accuracy across different memory sizes (data number per class).

When memory is limited (e.g., 50 images per class), the gap between DERPLL and random selection is at its widest, proving that a carefully curated summary of data is substantially more effective than a larger, unoptimized subset.

Conclusion

DERPLL provides a practical framework for maintaining secure and accurate models in lifelong learning settings. By focusing on "quality over quantity" through bi-level optimization, it achieves state-of-the-art results without the need for external unlabeled data. Future work will focus on reducing the computational overhead of the optimization process and scaling the approach to larger datasets like ImageNet.

Summary updated

Abstract

Lifelong learning (LL) aims to improve a predictive model as the data source evolves continuously. Most work in this learning paradigm has focused on resolving the problem of 'catastrophic forgetting,' which refers to a notorious dilemma between improving model accuracy over new data and retaining accuracy over previous data. Yet, it is also known that machine learning (ML) models can be vulnerable in the sense that tiny, adversarial input perturbations can deceive the models into producing erroneous predictions. This motivates the research objective of this paper - specification of a new LL framework that can salvage model robustness (against adversarial attacks) from catastrophic forgetting. Specifically, we propose a new memory-replay LL strategy that leverages modern bi-level optimization techniques to determine the 'coreset' of the current data (i.e., a small amount of data to be memorized) for ease of preserving adversarial robustness over time. We term the resulting LL framework 'Data-Efficient Robustness-Preserving LL' (DERPLL). The effectiveness of DERPLL is evaluated for class-incremental image classification using ResNet-18 over the CIFAR-10 dataset. Experimental results show that DERPLL outperforms the conventional coreset-guided LL baseline and achieves a substantial improvement in both standard accuracy and robust accuracy.
BibTeX
@inproceedings{jia2023robustness,
  author = {Jinghan Jia and Yihua Zhang and Dogyoon Song and Sijia Liu and Alfred O. Hero III},
  title = {{Robustness-preserving Lifelong Learning via Dataset Condensation}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023},
  doi = {10.1109/icassp49357.2023.10096756},
}
Figure