CSTAR: Towards Compact and Structured Deep Neural Networks with Adversarial Robustness

Authors: Huy Phan, Miao Yin, Yang Sui, Bo Yuan, Saman Zonouz

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Evaluations for various DNN models on different datasets demonstrate the effectiveness of CSTAR. Compared with the state-of-the-art robust structured pruning, CSTAR shows consistently better performance. For instance, when compressing Res Net-18 on CIFAR-10, CSTAR achieves up to 20.07% and 11.91% improvement for benign accuracy and robust accuracy, respectively.
Researcher Affiliation Academia Huy Phan1, Miao Yin1, Yang Sui1, Bo Yuan1, Saman Zonouz2 1Department of Electrical and Computer Engineering, Rutgers University 2Schools of Cybersecurity and Privacy, Georgia Institute of Technology
Pseudocode Yes Algorithm 1: Tucker-2 proj. for solving Eqn. 12; Algorithm 2: The overall procedure of CSTAR
Open Source Code No The paper does not provide a specific statement or link regarding the public availability of its source code.
Open Datasets Yes The models are compressed on CIFAR-10/100 and Image Net datasets with different compression ratios ranging from 2 to 64 .
Dataset Splits No The paper does not explicitly detail the use of a separate 'validation' dataset split or its specific proportion, only mentioning 'training' and 'testing' iterations.
Hardware Specification Yes Table 6 reports the inference time with batch size = 1 for the original and compressed models on both CPU (AMD Ryzen 9 5900HX) and GPU (NVIDIA Ge Force RTX 3090).
Software Dependencies No The paper mentions 'Py Torch and Tensor Flow platforms via using torch.tensordot and tf.tensordot' but does not specify their version numbers or other software dependencies with versions.
Experiment Setup Yes On CIFAR-10/100 L is selected with = 8/255, and PGD with step size α = 2/255 serves to generate adversarial examples. Here the number of PGD iterations for training and testing are 10 and 50, respectively. Other hyperparameter settings can be found in the Appendix.