Logo

Aref Mousavi

Powered by Jekyll — Theme by BDHU

Table of Contents

About Me

I am an M.Sc. student in Artificial Intelligence and a graduate researcher at the Robust and Interpretable Machine Learning (RIML) Lab at Sharif University of Technology, under the supervision of Prof. Mohammad Hossein Rohban. My research focuses on trustworthy machine learning, particularly adversarial robustness, model security, and robust representation learning, with a special interest in speech foundation models.

I also work directly with Prof. Mohammad Sabokrou as a researcher in his team, focusing on the trustworthiness and robustness of speech foundation models. Previously, I received my B.Sc. in Electrical Engineering from Isfahan University of Technology, where I conducted research under the supervision of Prof. Mohammad Reza Ahmadzadeh.

Education

Sharif University of Technology
M.Sc. in Artificial Intelligence (Computer Science & Engineering)
Sep 2025 - Present
Supervisor: Prof. Mohammad Hossein Rohban
GPA: 20.0/20.0 — Ranked 1st in the M.Sc. program


Isfahan University of Technology
B.Sc. in Electrical Engineering
Jul 2025
Supervisor: Prof. Mohammad Reza Ahmadzadeh
GPA: 18.58/20.0 (19.75/20.0 over the last two semesters) — Top 5% of B.Sc. graduates
Additionally, completed core Computer Engineering coursework.


Research Experience

RIML Lab
Graduate Researcher
Robust and Interpretable Machine Learning (RIML) Lab, Sharif University of Technology
Sep 2025 - Present
  • Researching adversarial robustness, representation learning, and backdoor vulnerabilities in speech foundation models under the supervision of Prof. Mohammad Hossein Rohban, and collaborating closely with Mojtaba Nafez (Idiap Research Institute; Ph.D. student, EPFL).
  • Work includes TransferBreaker (NeurIPS 2026, co-first author) and GhostWord (TMLR).


Researcher
Prof. Mohammad Sabokrou's Research Team (NewUU; Researcher, OIST)
Jun 2026 - Present
  • Working directly with Prof. Mohammad Sabokrou on adversarial robustness and representation learning for speech foundation models.
Joint Research: Led two jointly supervised research projects with Profs. Mohammad Hossein Rohban and Mohammad Sabokrou, resulting in first-author submissions to ICASSP 2027 and ICLR 2027.


Publications

TransferBreaker overview Breaking Adversarial Transferability in Fine-Tuned Speech Recognition

Mojtaba Nafez*, Aref Mousavi*, Mohammad Ebrahim Mahdavi, Mobina Poulaei, Kiarash Kiani Feriz, Mohammad Hossein Rohban.
NeurIPS 2026 (* equal contribution)

[Paper] [Code]

GhostWord overview GhostWord: A Fine-Grained Backdoor Attack on Automatic Speech Recognition

Mojtaba Nafez, Mobina Poulaei, Kiarash Kiani Feriz, Aref Mousavi, Mohammad Ebrahim Mahdavi, Mohammad Hossein Rohban.
TMLR

[Paper] [Code]

Hierarchical Stabilization overview Transferable Adversarial Robustness for Speech Foundation Models via Hierarchical Stabilization

Aref Mousavi, Shahab Sherafat, Kiarash Kiani Feriz, Amirparsa Safari, Raoof Zare Moayedi, Mohammad Hossein Rohban, Mohammad Sabokrou.
Submitted to ICASSP 2027

[Paper] [Code]

Teaching Experience

Teaching Assistant — Deep Learning
Sharif University of Technology
Sep 2026 - Present
Instructor: Prof. Mahdieh Soleymani


Teaching Assistant
Isfahan University of Technology
Feb 2025 - Jul 2025
Machine Learning — Instructor: Prof. Mehran Safayani
Designed and graded assignments, led problem-solving sessions, and evaluated final course projects.

Robotics — Instructor: Prof. Hamed Jalaly
Led problem-solving sessions, answered student questions, and graded assignments.


Selected Projects

VLM-Adversarial-Attack Thumbnail Exploring Adversarial Vulnerability of Vision-Language Models (VLMs)
Performed targeted FGSM, momentum PGD, and C&W attacks on images to manipulate the outputs of the BLIP image-to-caption model toward predefined captions, using random initialization, learning-rate decay, and hierarchical PGD-to-C&W initialization to improve attack performance while preserving visual quality.

EffiColorizer Thumbnail EffiColorizer: An Efficient Image Colorization Framework for Low-Power Devices
Developed an efficient image colorization framework using a spectrally normalized GAN with an EfficientNet-B3 encoder integrated into a U-Net generator and a hybrid training strategy, achieving a 2.4× improvement in FID while requiring approximately 10.8× fewer FLOPs than the baseline.

SmartFactoryLine Thumbnail SmartFactoryLine: Simulation-Based Industrial Automation and Control
Developed a simulated smart factory line using Siemens PLC control (TIA Portal), Factory I/O, and an OPC-connected Ignition HMI, integrating modular production-line automation, real-time monitoring, and closed-loop liquid-level regulation with alarms and automatic recovery.

QuadraArm Thumbnail QuadraArm: A 4-DOF Robotic Arm Platform from Hardware Design to Motion Planning
Developed a complete 4-DOF robotic arm platform spanning hardware design and analysis, forward and inverse kinematics, velocity kinematics, trajectory planning, and motion imitation via demonstration.

Awards & Honors