Mojtaba Nafez

I am a first-year Ph.D. student at École Polytechnique Fédérale de Lausanne (EPFL), Switzerland, and a Research Assistant in the Natural Language Understanding (NLU) group at the Idiap Research Institute.

I received my M.Sc. degree in Computer Engineering from Sharif University of Technology, where I conducted research in the Robust and Interpretable Machine Learning (RIML) Lab under the supervision of Prof. MohammadHossein Rohban , and collaborated closely with Dr. Mohammad Sabokrou .

Previously, I received my Bachelor of Science in Computer Engineering from Iran University of Science and Technology. In my final year, I worked as a research assistant at CVLab IUST under the supervision of Prof. Mohammad Reza Mohammadi .

Email  /  Google Scholar  /  Github  /  LinkedIn

I have extensive research experience in the trustworthiness and reliability of machine learning models, particularly in the domains of computer vision (image and video) and automatic speech recognition. I am now motivated to expand my work toward Large Language Models (LLMs).

Weakly Supervised Video Anomaly Detection (WSVAD) models are vulnerable to adversarial attacks, and standard defenses fail under weak supervision. We propose SRD, a pseudo-anomaly generation method that creates temporally consistent synthetic anomalies to reduce pseudo-label noise and enable effective adversarial training. Our approach improves robustness, outperforming prior methods by 71.0% AUROC on average across benchmarks.

PatchGuard is an adversarially robust Anomaly Detection (AD) and Localization (AL) method using pseudo anomalies in a ViT-based framework. It leverages Foreground-Aware Pseudo-Anomalies and a novel loss function for improved robustness, achieving significant performance gains on industrial and medical datasets under adversarial settings.

Anomaly Detection (AD) methods are vulnerable to adversarial attacks due to relying on unlabeled normal samples. The authors address this by creating a pseudo-anomaly group and using adversarial training with contrastive loss, mitigating spurious negative pairs through opposite pairs to improve robustness.

Addressing a critical practical challenge within the domain of image-based anomaly detection, our research confronts the absence of a universally applicable and adaptable methodology that can be tailored to diverse datasets characterized by distinct inductive biases.

In this research, our problem was to identify whether a given model was backdoored or not. The study finds that backdoored models exhibit jagged decision boundaries around out-of-distribution (OOD) samples, leading to reduced robustness.

Academic Ranking: 3rd among 105 students in Bachelor’s degree at Iran University of Science and Technology (IUST)

Design and source code from Jon Barron's website.