Detecting and Corrupting Convolution-based Unlearnable Examples
Authors: Minghui Li, Xianlong Wang, Zhifei Yu, Shengshan Hu, Ziqi Zhou, Longling Zhang, Leo Yu Zhang
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate the effectiveness of detection scheme EPD and that our defense COIN outperforms 11 state-of-the-art (SOTA) defenses, achieving a significant improvement on the CIFAR and Image Net datasets. |
| Researcher Affiliation | Academia | 1 School of Software Engineering, Huazhong University of Science and Technology 2 Hubei Key Laboratory of Distributed System Security, Hubei Engineering Research Center on Big Data Security 3 School of Cyber Science and Engineering, Huazhong University of Science and Technology 4 School of Computer Science and Technology, Huazhong University of Science and Technology 5 School of Information and Communication Technology, Griffith University |
| Pseudocode | No | The paper describes methods through mathematical equations and descriptive text, but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code https://github.com/wxldragon/COIN |
| Open Datasets | Yes | Benchmark datasets including CIFAR-10 (32 32), CIFAR-100 (32 32) (Krizhevsky and Hinton 2009), and Image Net (224 224) (Deng et al. 2009) are used. |
| Dataset Splits | No | Consistent with previous works (Wang et al. 2024a,c), we employ test accuracy, i.e., the model accuracy on a clean test set after training on the datasets. |
| Hardware Specification | Yes | Each result is run once on a 3090 GPU. |
| Software Dependencies | No | The paper mentions training with SGD and using various network architectures (Res Net, Dense Net, Mobile Net V2, VGG) but does not provide specific version numbers for any libraries, frameworks, or software used. |
| Experiment Setup | Yes | The uniform distribution range α of COIN is empirically set to 2.0. During pre-training the SVM classifier of EPD, the linear kernel function is utilized, and penalty parameter Cp is set to 0.1. Meanwhile, the threshold θ of EPD during detection is set to 0. During training of classifiers, we use SGD for training 80 epochs with a momentum of 0.9, a learning rate of 0.1, and batch sizes of 128, 32 for CIFAR and Image Net, respectively. |