LoRID: Low-Rank Iterative Diffusion for Adversarial Purification

Authors: Geigh Zollicoffer, Minh N. Vu, Ben Nebgen, Juan Castorena, Boian Alexandrov, Manish Bhattarai

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Consequently, Lo RID increases the effective diffusion time-steps and overcomes strong adversarial attacks, achieving superior robustness performance in CIFAR-10/100, Celeb A-HQ, and Image Net datasets under both white-box and grey-box settings. [...] Sect. 4 provides our experimental results, and Sect. 5 concludes this paper.
Researcher Affiliation Collaboration 1Georgia Institute of Technology, Atlanta, GA 2 Theoritical Division, Los Alamos National Laboratory, Los Alamos, NM 3 Computational Sciences, Los Alamos National Laboratory, Los Alamos, NM
Pseudocode Yes The pseudocode of Lo RID is described in Appendix. B.5.
Open Source Code No The paper does not provide an explicit statement about releasing source code for the methodology or a link to a code repository.
Open Datasets Yes We evaluate Lo RID on CIFAR-10/100 (Rabanser, Shchur, and G unnemann 2017), Celeb A-HQ (Karras et al. 2018), and Image Net (Deng et al. 2009).
Dataset Splits Yes When the gradients are not needed to pass through the defense (grey-box setting) in CIFAR-10, all methods are evaluated 10000 test images. On the other hand, due to the high computational cost of computing gradients for adaptive attacks against diffusion-based defenses, we assess the methods on a fixed subset of 512 randomly sampled test images, consistent with previous studies (Nie et al. 2022; Lee and Kim 2023).
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers like Python 3.8, CPLEX 12.4) needed to replicate the experiment.
Experiment Setup Yes Lo RID requires the specification of both the time-step t and the looping number L, which are crucial for its iterative process. These hyperparameters are generally selected by evaluating the classifier s performance on the clean dataset, with t and L chosen to maintain acceptable clean accuracy. Further details on this parameter selection process are provided in Appx. B.6. We report those parameters as a tuple (t, L) next to the name of our method. For example, in Table 2: "Lo RID (39, 5)" and "Lo RID (20, 24)".