I am a fifth-year Ph.D. candidate in the IS-WiN Lab at Clemson University (CU), advised by Dr. Fatemeh Afghah. Before joining CU, I worked as a research assistant on biometrics at West Virginia University. I completed my Master's degree in Electrical Engineering at Iran University of Science & Technology. My research focuses on enhancing the generalization of machine learning models to unseen domains, with applications spanning various areas, including anomaly detection, biometrics, healthcare, visual perception tasks, and scene understanding.
News
[2026/01] One paper on VLM-based anomaly detection has been accepted to ICASSP 2026.
[2025/05] One paper on Vision-Language Models has been accepted at KDD 2025.
[2025/02] FreqDebias has been accepted at CVPR 2025.
[2024/11] Two papers have been accepted at WACV 2025.
[2024/11] Successfully passed Qualifying Exam.
[2024/06] My new paper CATFace is accepted by IEEE Transactions on Biometrics, Behavior, and Identity Science.
[2024/02] AAFACE is accepted at IEEE ICIP 2023.
[2023/11] MedViT has been featured in Computers in Biology and Medicine.
[2023/09] Our new method on Morph Attack Detection has been accepted by IJCB 2023.
My current research focuses on advancing Multimodal Large Language Models to enhance the robustness and adaptability of anomaly detection in few-shot settings and provide detailed descriptions of each detected irregularity. My prior work at CU includes developing advanced prompting techniques for Vision-Language models and robust multi-class anomaly detection.
At WVU, in collaboration with CITER and NIST, I explored advanced security measures for automated face recognition systems, focusing on robust detection and generation of face morphing attacks. Additionally, I worked on multimodal biometric recognition under long-range constraints to address issues related to turbulent, low-quality imagery from extended distances.
I also engaged in advancing Vision Transformers for healthcare at IUST. This includes developing hybrid Transformers with locality inductive biases for medical imaging; enhancing adversarial robustness with innovative data augmentation techniques like Moment Exchanger and Patch Momentum Changer; improving local and global dependencies within Transformer architectures.
Beyond these areas, my research extends to visual perception and scene understanding, such as object tracking, detection, and segmentation. This includes enhancing the robustness, generalization, and adaptability of perception models such as object detection and semantic segmentation for autonomous vehicles; developing multi-teacher knowledge distillation framework to enhance the performance of lightweight perception models, ensuring they operate effectively in challenging environments.
We propose PromptMAD, a cross-modal prompting framework for multi-class visual anomaly localization that leverages vision-language models to improve anomaly detection and localization across multiple classes.
We introduce a novel frequency-domain debiasing framework that uses consistency regularization to improve the generalization of deepfake detectors to unseen manipulations.
We propose DiSa, a directional saliency-aware prompt learning framework that improves generalization in vision-language models by integrating cross-interactive regularization and directional regularization strategies.
We propose a robust multi-class anomaly detection framework with a class-aware prompt integration mechanism to mitigate inter-class interference and a domain adapter to handle domain shifts.
We propose a novel multi-branch network with cross-attribute-guided fusion and self-attention distillation, improving face recognition in low-quality images using soft biometric attributes.
This study proposes a robust and efficient CNN-Transformer hybrid model, combining CNN locality with the global connectivity of vision Transformers. Additionally, we enhance robustness by learning smoother decision boundaries through feature mean and variance permutation within mini-batches.
We propose consistency regularization to enhance the generalization of morph attack detection through morph-wise augmentations to enhance robustness against unseen morph attacks in biometric systems.
This paper proposes a multi-task neural network that generates a face quality vector, including nuisance factors, offering improved performance and detailed feedback for face image quality assessment.
This paper proposes a robust ensemble of CNNs and Transformers for morph detection that enhances generalization to morph attacks and increases robustness against adversarial threats through multi-perturbation training.
This paper proposes a robust transformer model that incorporates locality inductive bias and feature normalization, enhancing generalization and robustness in feature extraction tasks.
We propose a two-phase multi-expert classification method for human action recognition, addressing long-tailed distribution using super-class learning without extra data or manual annotation. A novel Graph-Based Class Selection (GCS) algorithm optimizes class configurations and inter-class dependencies.
We address the generalization issues in scene understanding for autonomous vehicles by employing GANs for weather modeling, and advanced augmentations, improving object detection robustness and generalization across domains, especially in adverse weather conditions and natural distortions.
This paper introduces a hybrid CNN-Transformer model for COVID-19 diagnosis using CT images, combining local and global feature extraction, and achieving superior performance with limited training data.
We propose a multi-teacher KD framework in which several expert CNNs, trained on different settings, supervise a lightweight student model. This framework enhances the robustness and performance of the student by using diverse knowledge sources.
This work aims to improve motion and observation models in visual object tracking. We propose a motion estimation network to refine target location predictions, with a Siamese network detecting the most probable candidate. Additionally, a weighting CNN adaptively assigns weights to similarity scores, accounting for target appearance changes.
