This work presents a novel reinforcement learning approach for generating universal adversarial suffixes for language models using calibrated reward mechanisms. The method demonstrates effectiveness in creating adversarial attacks while maintaining computational efficiency.
This work introduces a calibrated Gumbel-Softmax relaxation technique for generating universal adversarial suffixes. The proposed method provides a differentiable approach to discrete token optimization in language model attacks.
This work introduces a novel contrastive learning approach for enhancing adversarial robustness in deep neural networks. We demonstrate significant improvements in defense against various attack methods while maintaining competitive clean accuracy.
A Pramanick, M Bansal, U Srivastava, Suklav Ghosh, Arijit Sur
arXiv preprint arXiv:2510.27245, 2025
This work presents a novel transformer-based denoising approach for adversarial defense that operates in both spatial and frequency domains. The method effectively mitigates adversarial perturbations while preserving essential image features.
Multi-source Transfer Learning & Self-Supervised Learning for Enhanced Visual Representation
In Progress
Suklav Ghosh, Arijit Sur, Pinaki Mitra
Under Preparation, 2025
An investigation into combining multiple source domains for transfer learning with self-supervised learning techniques. The work aims to improve visual representation learning across diverse domains and tasks.
International Conference on Innovations in Data Analytics (ICIDA), Springer Nature Singapore, 2023
This paper presents a novel non-recursive, space-efficient algorithm for solving the N-Queens problem. The proposed blind approach efficiently finds all possible solutions while maintaining optimal space complexity, making it suitable for large-scale applications.
A Model-Based Approach For Flight Delay Prediction
Published
A Dutta, M Nandi, S Sarkar, S Hota, Suklav Ghosh
Indian Journal of Applied Research, 2023
A comprehensive study on flight delay prediction using machine learning models. The research explores various factors contributing to flight delays and develops predictive models to assist in aviation management and passenger planning.
Journal Article
Research Focus Areas
Multimodal Learning
Vision-language models and cross-modal understanding systems
Computer Vision
Image analysis, object detection, and visual understanding systems
Self-Supervised Learning
Learning representations from unlabeled data using contrastive methods
Deep Learning
Neural networks, representation learning, and model optimization techniques
Future Directions
My current research focuses on advancing the robustness and efficiency of deep learning models through novel self-supervised learning approaches and multi-modal understanding. I am particularly interested in developing methods that can learn from minimal supervision while maintaining strong performance across diverse domains.
Multimodal LearningFew-Shot LearningDomain AdaptationRobust AI