📝 Selected Publications

On the Robustness of Medical Vision-Language Models: Are they Truly Generalizable?
Raza Imam^#, Rufael Marew, Mohammad Yaqub. (MIUA 2025) code

• TLDR;

  We introduce MediMeta-C (with MedMNIST-C) to benchmark MVLM robustness under real-world corruptions and propose RobustMedCLIP, a lightweight few-shot, low-rank adaptation that significantly restores performance on noisy medical images without sacrificing accuracy on clean data.

From RAG to Agentic: Validating Islamic-Medicine Responses with LLM Agents
Mohammad Amaan Sayeed^*, Mohammed Talha Alam^*, Raza Imam^* Shahab Saquib Sohail, Amir Hussain (ICMLxMusiML 2025)

• TLDR;

  We introduce TibbeAG, a pipeline that benchmarks LLaMA-3, Mistral7B, and Qwen2-7B on Prophetic Medicine QA by combining retrieval-augmented generation and agentic self-critique—yielding a 13% boost in factual accuracy and a further 10% gain in mechanistic insight and safety.

Noise is an efficient learner for zero-shot vision-language models
Raza Imam^#, Asif Hanif, Jian Zhang, Khaled Waleed Dawoud, Yova Kementchedjhieva, Mohammad Yaqub. (ICCVxCLVL 2025) code

• TLDR;

  We introduce Test-Time Noise Tuning (TNT), which learns input-space noise and enforces multi-view embedding alignment to adapt vision-language models on unlabeled test samples, yielding +7.38% on natural distribution shifts and +0.80% on cross-dataset benchmarks over zero-shot CLIP.

Test-Time Low Rank Adaptation via Confidence Maximization for Zero-Shot Generalization of Vision-Language Models
Raza Imam^#, Hanan Ghani, Muhammad Huzaifa, Karthik Nandakumar. (WACV 2025) code

• TLDR;

  Introduces Test-Time Low-rank Adaptation (TTL) as a parameter-efficient alternative to prompt tuning for zero-shot generalization of VLMs. TTL updates transformer attention weights during inference by maximizing prediction confidence using a weighted entropy loss for consistency across augmented samples.

SEDA: Self-ensembling ViT with Defensive Distillation and Adversarial Training for Robust Chest X-Rays Classification
Raza Imam^#, Ibrahim Almakky, Salma Alrashdi, Baketah Alrashdi, Mohammad Yaqub. (MICCAI 2023) code

• TLDR;

  SEDA utilizes efficient CNN blocks to learn spatial features with various levels of abstraction from feature representations extracted from intermediate ViT blocks, that are largely unaffected by adversarial perturbations. SEDA achieves computational efficiency of 70x times lighter framework and enhanced robustness of +9%.

On enhancing the robustness of Vision Transformers: Defensive Diffusion
Raza Imam, Muhammad Huzaifa^#, Mohammed El-Amine Azz. (MIUA 2023) code

• TLDR;

  We introduced a defensive diffusion technique as an adversarial purifier to eliminate adversarial noise introduced by attackers in the original image. Extensive experiments conducted on a publicly available Tuberculosis X-ray dataset validate the computational efficiency and improved robustness achieved by our proposed architecture.

A systematic literature review of attribute based encryption in health services
Raza Imam, Kaushal Kumar, Syed Mehran Raza, Rumi Sadaf, Faisal Anwer^#, Noor Fatima, Mohammad Nadeem, Mohamed Abbas, Obaidur Rahman. (Elsevier JKSU)

• TLDR;

  A systematic and comprehensive study of ABE works concerning E-Health as the authors rigorously investigate healthcare-focused ABE frameworks. This will help future research in ABE to secure the existing E-Health data sharing more effectively.