Article Per Year (5 Year)
p-Index From 2021 - 2026
0.444
P-Index
This Author published in this journals
All Journal International Journal of Advanced Science Computing and Engineering
Sharma, Shorya
Unknown Affiliation
Computer Science & IT Control & Systems Engineering Electrical & Electronics Engineering

Published : 2 Documents Claim Missing Document
Claim Missing Document
Check
Articles

Found 2 Documents
Search

 1 
Multi-SAP Adversarial Defense for Deep Neural Networks Sharma, Shorya
International Journal of Advanced Science Computing and Engineering Vol. 4 No. 1 (2022)
Publisher : SOTVI

Show Abstract | Download Original | Original Source | Check in Google Scholar | Full PDF (500.333 KB) | DOI: 10.62527/ijasce.4.1.76

Abstract

Deep learning models have gained immense popularity for machine learning tasks such as image classification and natural language processing due to their high expressibility. However, they are vulnerable to adversarial samples - perturbed samples that are imperceptible to a human, but can cause the deep learning model to give incorrect predictions with a high confidence. This limitation has been a major deterrent in the deployment of deep learning algorithms in production, specifically in security critical systems. In this project, we followed a game theoretic approach to implement a novel defense strategy, that combines multiple Stochastic Activation Pruning with adversarial training. Our defense accuracy outperforms that of PGD adversarial training, which is known to be the one of the best defenses against several L∞ attacks, by about 6-7%. We are hopeful that our defense strategy can withstand strong attacks leading to more robust deep neural network models.Deep learning models have gained immense popularity for machine learning tasks such as image classification and natural language processing due to their high expressibility. However, they are vulnerable to adversarial samples - perturbed samples that are imperceptible to a human, but can cause the deep learning model to give incorrect predictions with a high confidence. This limitation has been a major deterrent in the deployment of deep learning algorithms in production, specifically in security critical systems. In this project, we followed a game theoretic approach to implement a novel defense strategy, that combines multiple Stochastic Activation Pruning with adversarial training. Our defense accuracy outperforms that of PGD adversarial training, which is known to be the one of the best defenses against several L∞ attacks, by about 6-7%. We are hopeful that our defense strategy can withstand strong attacks leading to more robust deep neural network models.
Generating Accurate Human Face Sketches from Text Descriptions Sharma, Shorya
International Journal of Advanced Science Computing and Engineering Vol. 6 No. 1 (2024)
Publisher : SOTVI

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.62527/ijasce.6.1.195

Abstract

Drawing a face for a suspect just based on the descriptions of the eyewitnesses is a difficult task. There are some state-of-the-art methods in generating images from text, but there is only a little research in generating face images from text, and almost none in generating sketches from text. As a result, there is no dataset available to tackle this task.  We developed a text-to-sketch dataset derived from the CelebA dataset, which comprises 200,000 celebrity images, thereby facilitating the investigation of the novel task of generating police sketches from textual descriptions. Furthermore, we demonstrated that the application of AttnGAN for generating sketch images effectively captures the facial features described in the text. We identified the optimal configuration for AttnGAN and its variants through experiments involving various recurrent neural network types and embedding sizes. We provided commonly used metric values, such as the Inception score and Fréchet Inception Distance (FID), for the two-attention-based state-of-the-art model we achieved. However, we also identified areas for improvement in the model's application. Experiments conducted with a new dataset consisting of 200 sketch images from Beijing Normal University revealed that the model encounters challenges when handling longer sentences or unfamiliar terms within descriptions. This limitation in capturing features from such text contributes to a decrease in image diversity and realism, adversely impacting the overall performance of the model. For future improvements, consider exploring alternative models such as Stack-GAN, Conditional-GAN, DC-GAN, and Style-GAN, which are known for their capabilities in face image generation. Simplifying architecture while maintaining performance can also help deploy models on mobile devices for real-world use.
Co-Authors