Publications
A Unified, Resilient, and Explainable Adversarial Patch Detector
Vishesh Kumar, Akshay Agarwal
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
Abstract: The AdvPatchXAI is a defense algorithm designed to protect Deep Neural Networks (DNNs) from physical adversarial attacks. It uses a novel patch decorrelation loss, self-supervised learning, and a sparse linear layer for classification. The model improves interpretability and correlation with human vision, closing the semantic gap between latent and pixel space and effectively handling unseen patches, thereby enhancing DNN robustness in practical settings.
Detection of identity swapping attacks in low-resolution image settings
Akshay Agarwal, Nalini Ratha
Journal of Information Security and Applications
Abstract: This research proposes a low-resolution identity swap attack detection algorithm to address the challenge of detecting fake face images in low-resolution videos. The algorithm uses artifacts amplification and classification to handle the lack of information content. Extensive evaluations using multiple databases, resolution settings, and attack types demonstrate the strength and effectiveness of the proposed algorithm in in-the-wild settings. The results show superiority compared to existing state-of-the-art works.
On Which Data Distribution (Synthetic or Real) We Should Rely for Soft Biometric Classification
Manju RA, Atul Kumar, Akshay Agarwal
Winter Conference on Applications of Computer Vision (WACV)
Abstract: Gender identification is vital for human-computer interaction and identity search. While "real" facial images are standard in gender classification, synthetic images are gaining attention due to privacy concerns and advancements in generative networks. However, their effectiveness remains unclear. This study evaluates the performance of gender classification networks trained on real vs. synthetic face images. Using 8 DNNs, including CNNs and ViTs, across 4 datasets and 6 image corruptions, we also employ Grad-CAM and t-SNE for interpretability.
PrecipFormer: Efficient Transformer for Precipitation Downscaling
R. Kumar, T. Sharma, V. Vaghela, S. Jha, A. Agarwal
Winter Conference on Applications of Computer Vision (WACV) Workshops
Abstract: Precipitation downscaling is a crucial challenge in climate modeling and hydrological applications. This paper introduces PrecipFormer, a computationally efficient transformer architecture designed for this task. It builds upon the Low-to-High Multi-Level Vision Transformer mechanism, enabling parallel processing of features at multiple spatial scales and reducing computational overhead. PrecipFormer achieves superior performance compared to state-of-the-art baselines.
