CausalDiff: Causality-Inspired Disentanglement via Diffusion Model for Adversarial Defense

Authors: Mingkun Zhang, Keping Bi, Wei Chen, Quanrun Chen, Jiafeng Guo, Xueqi Cheng

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, Causal Diff has significantly outperformed state-of-the-art defense methods on various unseen attacks, achieving an average robustness of 86.39% (+4.01%) on CIFAR-10, 56.25% (+3.13%) on CIFAR-100, and 82.62% (+4.93%) on GTSRB (German Traffic Sign Recognition Benchmark).
Researcher Affiliation Academia Mingkun Zhang CAS Key Laboratory of AI Safety Institute of Computing Technology, CAS zhangmingkun20z@ict.ac.cn Keping Bi Key Laboratory of Network Data Science and Technology Institute of Computing Technology, CAS bikeping@ict.ac.cn Wei Chen CAS Key Laboratory of AI Safety Institute of Computing Technology, CAS chenwei2022@ict.ac.cn Quanrun Chen School of Statistics University of International Business and Economics qchen@uibe.edu.cn Jiafeng Guo Key Laboratory of Network Data Science and Technology Institute of Computing Technology, CAS guojiafeng@ict.ac.cn Xueqi Cheng CAS Key Laboratory of AI Safety Institute of Computing Technology, CAS cxq@ict.ac.cn
Pseudocode Yes Algorithm 1 Causal Diff Algorithm; Algorithm 2 Causal Diff Pretrain Algorithm; Algorithm 3 Adversarially Robust Inference Algorithm
Open Source Code Yes The code is available at https://github.com/CAS-AISafety Basic Research Group/Causal Diff.
Open Datasets Yes Our experiments utilize the CIFAR-10, CIFAR-100 [18] and GTSRB [19] datasets.
Dataset Splits No The paper mentions 'CIFAR-10 and CIFAR-100 each consists of 50,000 training images, categorized into 10 and 100 classes, respectively.' and 'GTSRB comprises 39,209 training images' but does not explicitly state the train/validation/test splits, percentages, or specific counts for these datasets within the paper.
Hardware Specification Yes We evaluate the computational complexity of Causal Diff and Diff Pure [33] as well as a discriminative model (WRN-70-16) by measuring the inference time in seconds for a single sample (average on 100 examples from CIFAR-10 dataset) on two types of GPUs, including NVIDIA A6000 GPU and 4090 GPU (Our experiments leverage 4 A6000 GPUs and 4 4090 GPUs).
Software Dependencies No The paper mentions using 'DDPM [17]' and 'Wide Res Net-70-16 (WRN-70-16)' and the 'Adam optimizer', but does not specify version numbers for these or other software dependencies like Python, PyTorch, or CUDA.
Experiment Setup Yes Both the pretraining and joint training phases utilize a learning rate of 1e 4 and a batch size of 128. For simplicity, we follow the setting of wt = 1 [17]. We set α = 1., γ = 1e 2, η = 1e 5, λ = 1e 2 as the weights for the loss function in Eq. (7).