Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Enhancing Robustness in Incremental Learning with Adversarial Training
Authors: Seungju Cho, Hongsin Lee, Changick Kim
AAAI 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that FLAIR significantly outperforms other baselines. To the best of our knowledge, we are the first to comprehensively investigate the baselines, challenges, and solutions for ARCIL, which we believe represents a significant advance toward achieving real-world robustness. |
| Researcher Affiliation | Academia | Korea Advanced Institute of Science and Technology (KAIST) EMAIL |
| Pseudocode | No | The paper includes mathematical formulations of the methods, such as the loss functions for different adversarial training and incremental learning approaches in Table 1, and the FLAIR loss function, but it does not present any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code can be accessed at https://github.com/Hongsin Lee/FLAIR. |
| Open Datasets | Yes | Split CIFAR-10 (S-CIFAR10) divides CIFAR-10 (Krizhevsky 2009) into five tasks, each consisting of two classes. Split CIFAR-100 (S-CIFAR100) divides CIFAR-100 (Krizhevsky 2009) into ten tasks, with ten classes per task. Split SVHN (S-SVHN) divides the SVHN (Netzer et al. 2011) into five tasks, with two classes per task. Additionally, Split Tiny Image Net (STiny Image Net), which divides Tiny Image Net (Le and Yang 2015) into ten tasks with 20 classes per task using a memory buffer, can be found in the appendix. |
| Dataset Splits | Yes | Split CIFAR-10 (S-CIFAR10) divides CIFAR-10 (Krizhevsky 2009) into five tasks, each consisting of two classes. Split CIFAR-100 (S-CIFAR100) divides CIFAR-100 (Krizhevsky 2009) into ten tasks, with ten classes per task. Split SVHN (S-SVHN) divides the SVHN (Netzer et al. 2011) into five tasks, with two classes per task. |
| Hardware Specification | No | The paper mentions using the ResNet-18 architecture and MobileNet V2 for experiments, and parameters for PGD attacks, but it does not specify any particular hardware (e.g., GPU models, CPU types) used for running the experiments. |
| Software Dependencies | No | The paper mentions various methods and architectures like PGD-AT, TRADES, MART, ResNet-18, MobileNet V2, Rand Augment (RA), Auto Augment (Au A), but does not provide specific version numbers for any software libraries, frameworks, or programming languages used (e.g., Python, PyTorch, TensorFlow, CUDA versions). |
| Experiment Setup | Yes | For adversarial training, we used ten steps of PGD attack with a random start, and the maximum perturbation is limited to ϵ = 8/255, while each step is taken with a step size of 2/255. In all experiments, we utilized the Res Net-18 architecture (He et al. 2016), with additional results on Mobile Net V2 (Sandler et al. 2018) provided in the appendix. We conducted a grid search over the hyperparameters {0,0.5,1,2,4} for our method and all baseline methods, reporting the best results. |