Maximization of Average Precision for Deep Learning with Adversarial Ranking Robustness

Authors: Gang Li, Wei Tong, Tianbao Yang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical studies, which compare our method to current leading adversarial training baselines and other robust AP maximization strategies, demonstrate the effectiveness of the proposed approach. Notably, our methods outperform a state-of-the-art method (TRADES) by more than 4% in terms of robust AP against PGD attacks while achieving 7% higher AP on clean data simultaneously on CIFAR10 and CIFAR100.
Researcher Affiliation Collaboration Gang Li Texas A&M University College Station, USA gang-li@tamu.edu Wei Tong General Motors Warren, USA wei.tong@gm.com Tianbao Yang Texas A&M University College Station, USA tianbao-yang@tamu.edu
Pseudocode Yes Algorithm 1 Stochastic Algorithm for Solving Ad AP_LN in (9)
Open Source Code Yes The code is available at: https://github.com/Gang Lii/Adversarial-AP
Open Datasets Yes Datasets. We conduct experiments on four distinct datasets sourced from various domains. These encompass CIFAR-10 and CIFAR-100 datasets [22], Celeb A dataset [26] and the BDD100K dataset [57].
Dataset Splits Yes We split the training dataset into train/validation sets at 80%/20% ratio, and use the testing dataset for testing.
Hardware Specification Yes All experiments in our paper are run across 16 NVIDIA A10 GPUs and 10 NVIDIA A30 GPUs.
Software Dependencies No The paper mentions using ResNet18 as a backbone network and the Adam optimizer, but it does not provide specific version numbers for any software libraries, programming languages, or development environments used for the experiments.
Experiment Setup Yes We employ the Res Net18 [19] as the backbone network in our experiments. [...] For all methods, with mini-batch size as 128, we tune learning rate in {1e-3,1e-4,1e-5} with standard Adam optimizer. We set the weight decay to 2e-4 for the CIFAR10 and CIFAR100 datasets and 1e-5 for the Celeb A and BDD100k datasets. In the case of the CIFAR10 and CIFAR100 datasets, we run each method for a total of 60 epochs. For the Celeb A and BDD100k datasets, we run each method for 32 epochs. The learning rate decay happens at 50% and 75% epochs by a factor of 10. For MART, Ad AP_LN and Ad AP_LPN, we tune the regularization parameter λ in {0.1, 0.4, 0.8, 1, 4, 8, 10}. For TRADES, we tune the regularization parameter in {1, 4, 8, 10, 40, 80, 100}, since they favor larger weights to obtain better robustness. In addition, for Ad AP_LN and Ad AP_LPN, we tune its moving average parameters γ1, γ2 in {0.1, 0.9}. Similarly, we tune the moving average parameters γ1 for AP maximization in {0.1, 0.9}. We set margin parameter in the surrogate loss of AP as 0.6 for all methods that use the AP surrogate loss. For all adversarial training methods, we apply 6 projected gradient ascent steps to generate adversarial samples in the training stage, and the step size is 0.01. We choose L norm to bound the perturbation within the limit of ϵ = 8/255, as it is commonly used in the literature.