Splitting & Integrating: Out-of-Distribution Detection via Adversarial Gradient Attribution
Jiayu Zhang, Xinyi Wang, Zhibo Jin, Zhiyu Zhu, Jianlong Zhou, Fang Chen, Huaming Chen
Abstract
Out-of-distribution (OOD) detection is essential for enhancing the robustness and security of deep learning models in unknown and dynamic data environments. Gradient-based OOD detection methods, such as GAIA, analyse the explanation pattern representations of in-distribution (ID) and OOD samples by examining the sensitivity of model outputs w.r.t. model inputs, resulting in superior performance compared to traditional OOD detection methods. However, we argue that the non-zero gradient behaviors of OOD samples do not exhibit significant distinguishability, especially when ID samples are perturbed by random perturbations in high-dimensional spaces, which negatively impacts the accuracy of OOD detection. In this paper, we propose a novel OOD detection method called S & I based on layer Splitting and gradient Integration via Adversarial Gradient Attribution. Specifically, our approach involves splitting the model's intermediate layers and iteratively updating adversarial examples layer-by-layer. We then integrate the attribution gradients from each intermediate layer along the attribution path from adversarial examples to the actual input, yielding true explanation pattern representations for both ID and OOD samples. Experiments demonstrate that our S & I algorithm achieves state-of-the-art results, with the average FPR95 of 29.05% (ResNet34)/38.61% (WRN40) and 37.31% (BiT-S) on the CIFAR100 and Im-ageNet benchmarks, respectively. Our code is
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
1 top-tier paper has cited it since. Know what they found before your reviewers do.
Lune can draw on
the full paper · the top-tier papers that cite it · related work
Free to start. No credit card required.
Lineage
- Scaling for Training Time and Post-hoc Out-of-distribution Detection Enhancement cites Energy-based Out-of-distribution Detection
- Scaling for Training Time and Post-hoc Out-of-distribution Detection Enhancement cites ReAct: Out-of-distribution Detection With Rectified Activations
- On the Importance of Gradients for Detecting Distributional Shifts in the Wild cites Energy-based Out-of-distribution Detection
- ReAct: Out-of-distribution Detection With Rectified Activations cites Energy-based Out-of-distribution Detection
- ReAct: Out-of-distribution Detection With Rectified Activations cites On the Importance of Gradients for Detecting Distributional Shifts in the Wild
Cited by top-tier papers1
Builds on9
- Energy-based Out-of-distribution DetectionWeitang Liu, Xiaoyun Wang, John D. Owens, Yixuan LiNeurIPS 2020 · 2,213 citations
- ReAct: Out-of-distribution Detection With Rectified ActivationsYiyou Sun, Chuan Guo, Yixuan LiNeurIPS 2021 · 733 citations
- On the Importance of Gradients for Detecting Distributional Shifts in the WildRui Huang, Andrew Geng, Yixuan LiNeurIPS 2021 · 515 citations
- Detecting Out-of-Distribution Examples with Gram MatricesChandramouli Shama Sastry, Sageev OoreICML 2020 · 275 citations
- Scaling for Training Time and Post-hoc Out-of-distribution Detection EnhancementKai Xu, Rongyu Chen, Gianni Franchi, Angela YaoICLR 2024 · 81 citations
Related papers
- GAIA: Delving into Gradient-based Attribution Abnormality for Out-of-distribution DetectionJinggang Chen, Junjie Li, Xiaoyang Qu, Jianzong Wang et al.NeurIPS 2023 · 17 citations
- WeiPer: OOD Detection using Weight Perturbations of Class ProjectionsMaximilian Granz, Manuel Heurich, Tim LandgrafNeurIPS 2024 · 6 citations
- Iterative Search Attribution for Deep Neural NetworksZhiyu Zhu, Huaming Chen, Xinyi Wang, Jiayu Zhang et al.ICML 2024 · 5 citations
- Distilled Gradient Aggregation: Purify Features for Input Attribution in the Deep Neural NetworkGiyoung Jeon, Haedong Jeong, Jaesik ChoiNeurIPS 2022 · 11 citations
- Leveraging Perturbation Robustness to Enhance Out-of-Distribution DetectionWenxi Chen, Raymond A. Yeh, Shaoshuai Mou, Yan GuCVPR 2025
