Attributional Robustness Training Using Input-Gradient Spatial Alignment

Singh, Mayank; Kumari, Nupur; Mangla, Puneet; Sinha, Abhishek; Balasubramanian, Vineeth N; Krishnamurthy, Balaji

doi:10.1007/978-3-030-58583-9_31

Cited by 9 publications

(25 citation statements)

References 30 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…CAM [51] 62.57 ART [36] 75.45 Ours (method I) 76.30 Ours (method II) 76.70 Table 4. Results for CUB [41] segmentation.…”

Section: Methods Pxapmentioning

confidence: 99%

See 1 more Smart Citation

Learning a Weight Map for Weakly-Supervised Localization

Shaharabany¹,

Wolf²

2021

Preprint

View full text Add to dashboard Cite

In the weakly supervised localization setting, supervision is given as an image-level label. We propose to employ an image classifier f and to train a generative network g that outputs, given the input image, a per-pixel weight map that indicates the location of the object within the image. Network g is trained by minimizing the discrepancy between the output of the classifier f on the original image and its output given the same image weighted by the output of g. The scheme requires a regularization term that ensures that g does not provide a uniform weight, and an early stopping criterion in order to prevent g from over-segmenting the image. Our results indicate that the method outperforms existing localization methods by a sizable margin on the challenging fine-grained classification datasets, as well as a generic image recognition dataset. Additionally, the obtained weight map is also state-of-the-art in weakly supervised segmentation in fine-grained categorization datasets.

show abstract

“…CAM [51] 62.57 ART [36] 75.45 Ours (method I) 76.30 Ours (method II) 76.70 Table 4. Results for CUB [41] segmentation.…”

Section: Methods Pxapmentioning

confidence: 99%

“…For augmentation, the training procedure em-Figure 7. CUB weakly supervised segmentation where the first row is the input image, the second row is the ground-truth mask, the third row is the output map of [36] and the last row is our output map M . ploys a resize to a fixed size followed by a random crop, as well as a random horizontal flip.…”

Section: Methodsmentioning

confidence: 99%

Learning a Weight Map for Weakly-Supervised Localization

Shaharabany¹,

Wolf²

2021

Preprint

View full text Add to dashboard Cite

show abstract

“…Moreover, to ensure fair comparisons, we keep the hyper-parameters the same for models with or without IGR. (Chen et al, 2019) 36.13% 0.1562 51.84% 0.3446 74.49% 0.5811 IG-SUM-NORM (Chen et al, 2019) 41.53% 0.2240 57.27% 0.4097 78.70% 0.6901 AdvAAT (Ivankay et al, 2020) 51.74% 0.3791 73.62% 0.5810 72.11% 0.5484 ART (Singh et al, 2020b) 30.38% 0.1439 31.71% 0.2079 70.44% 0.6875 SSR (Wang et al, 2020) 38…”

Section: Experimental Configurationsmentioning

confidence: 99%

“…To compare, attribution protection methods, IG-NORM and IG-SUM-NORM by Chen et al (2019), Smooth Surface Regularization (SSR) (Wang et al, 2020), Attributional Robustness Traning (ART) (Singh et al, 2020b) and Adversarial Attributional Training with robust training loss (AdvAAT), are implemented and evaluated on all the datasets. A cross-entropy loss trained natural model (standard) is also included as a baseline.…”

Section: Evaluation On Attribution Robustnessmentioning

confidence: 99%

Exploiting the Relationship Between Kendall's Rank Correlation and Cosine Similarity for Attribution Protection

Wang¹,

Kong²

2022

Preprint

View full text Add to dashboard Cite

Model attributions are important in deep neural networks as they aid practitioners in understanding the models, but recent studies reveal that attributions can be easily perturbed by adding imperceptible noise to the input. The non-differentiable Kendall's rank correlation is a key performance index for attribution protection. In this paper, we first show that the expected Kendall's rank correlation is positively correlated to cosine similarity and then indicate that the direction of attribution is the key to attribution robustness. Based on these findings, we explore the vector space of attribution to explain the shortcomings of attribution defense methods using p norm and propose integrated gradient regularizer (IGR), which maximizes the cosine similarity between natural and perturbed attributions. Our analysis further exposes that IGR encourages neurons with the same activation states for natural samples and the corresponding perturbed samples, which is shown to induce robustness to gradient-based attribution methods. Our experiments on different models and datasets confirm our analysis on attribution protection and demonstrate a decent improvement in adversarial robustness.Preprint. Under review.

show abstract

“…Bae et al [29] used percentile to avoid the influence of extreme value when obtaining the final object location information. Singh et al [30] obtained more accurate contour of an object by antagonism confrontation enhancement. The method we proposed also has no objects proposal or additional batch for classification compared with others.…”

Section: Related Workmentioning

confidence: 99%

Category‐wise feature extractor based on ADL method for weak‐supervised object localisation

Zhu

et al. 2020

IET Image Processing

View full text Add to dashboard Cite

Fully supervised object detection needs to use the data set with object category annotation and location annotation to train the model. In contrast, weak supervised object localisation only needs to use the data set with object category annotation to train the model, but it can complete the classification task and object localisation task at the same time. Inspired by the attention-based dropout layer (ADL) method, this study designs a category-wise feature extractor (CFE), which can explicitly obtain the localisation map used to indicate the object location, and it is directly related to the category output of the classification task. Although its amount of calculation is slightly larger than that of the ADL method, the performance is better than ADL in some tasks. In addition to the standard CFE method mentioned above, this study also designs a lightweight CFEtiny method, which adopts split-attention mechanism, and the calculation amount of this method is much smaller than that of ADL method.

show abstract

Attributional Robustness Training Using Input-Gradient Spatial Alignment

Cited by 9 publications

References 30 publications

Learning a Weight Map for Weakly-Supervised Localization

Learning a Weight Map for Weakly-Supervised Localization

Exploiting the Relationship Between Kendall's Rank Correlation and Cosine Similarity for Attribution Protection

Category‐wise feature extractor based on ADL method for weak‐supervised object localisation

Contact Info

Product

Resources

About