RODEO: Robust Outlier Detection via Exposing Adaptive Out-of-Distribution Samples
Authors: Hossein Mirzaei, Mohammad Jafari, Hamid Reza Dehbashi, Ali Ansari, Sepehr Ghobadi, Masoud Hadi, Arshia Soltani Moakhar, Mohammad Azizmalayeri, Mahdieh Soleymani Baghshah, Mohammad Hossein Rohban
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate RODEO in both clean and adversarial settings. In the adversarial setting, numerous strong attacks, including PGD-1000 (Madry et al., 2017), Auto Attack (Croce & Hein, 2020), and Adaptive Auto Attack (A3) (Liu et al., 2022), are employed for robustness evaluation. Our experiments are conducted across various common outlier detection setups, including Novelty Detection (ND), Open-Set Recognition (OSR), and Out-of-Distribution (OOD) detection. |
| Researcher Affiliation | Academia | 1Sharif University of Technology, Tehran, Iran 2Isfahan University of Technology, Isfahan, Iran. |
| Pseudocode | Yes | Algorithm 1 RODEO: Adversarial Training with Adaptive Exposure Dataset |
| Open Source Code | Yes | For access to the implementation, please visit the following link: https://rohban-lab.github.io/rodeo. |
| Open Datasets | Yes | For the low-resolution datasets, we included CIFAR10 (Krizhevsky et al., 2009), CIFAR100 (Krizhevsky et al., 2009), MNIST (Le Cun & Cortes, 2010), and Fashion MNIST (Xiao et al., 2017). Furthermore, we performed experiments on medical and industrial high-resolution datasets, namely Head-CT (Kitamura, 2018), MVTec-ad (Bergmann et al., 2019), Brain-MRI (Bhuvaji et al., 2020), Covid19 (Cohen et al., 2020), and Tumor Detection (Nickparvar, 2021). |
| Dataset Splits | Yes | Open-Set Recognition. For this task, each dataset was randomly split into an inlier set and an outlier set at a 60/40 ratio. This random splitting was repeated 5 times. |
| Hardware Specification | Yes | Experiments were conducted on RTX 3090 GPUs. |
| Software Dependencies | No | The paper mentions using Res Net-18, Wide Res Net, and Adam optimizer, citing their original papers. However, it does not provide specific version numbers for any software libraries or frameworks used (e.g., Python, PyTorch, TensorFlow, scikit-learn versions). |
| Experiment Setup | Yes | Furthermore, RODEO is trained 100 epochs with Adam(Kingma & Ba, 2017) optimizer with a learning rate of 0.001 for each experiment. We set the value of ϵ to 8 255 for low-resolution datasets and 2 255 for high-resolution ones. For the PGD attack, we use 10 random restarts for the attack, with random initializations within the range of ( ϵ, ϵ), and perform N = 1000 steps. Furthermore, we select the attack step size as ³ = 2.5 ϵ N . |