Milad Bafarassat

Sabanci University

Developing CORTEX, a transformer-based beam selection system for terahertz (THz) UAV communications. The architecture combines convolutional front-ends with causal transformers using rotary position embeddings (RoPE) for drone state prediction. Key techniques include a two-stage approach for angle prediction and beam selection, circular-aware angle refinement, knowledge distillation, and focal loss for class imbalance. Paper accepted to the IEEE Open Journal of the Communications Society.

Supervised by Professor Ozgur Gurbuz

Sabanci University

Research focuses on security evaluation and attack development for state-of-the-art time-series models. Implemented and extended two recent approaches: (1) input-aware poisoning attacks that craft sample-specific triggers, and (2) generative-model based trigger synthesis (GAN/conditional generative models) to produce stealthy backdoors against RNNs, LSTMs and transformer-based time- series classifiers. Conducted threat-model design, empirical evaluations on standard benchmarks, and initial defenses/mitigations.

Supervised by Professor Emre Ozfatura

Shanghai AI Lab (Remote)

Researched mechanistic analysis and interpretability of transformer-based language models to enhance AI reliability and safety. Implemented a Meta-CoT-based approach to automate mechanistic interpretability in transformer architectures.

Supervised by Wenyu Du (HKU) and Dr. Jie Fu (Shanghai AI Lab)

Sabanci University (Graduation Project)

Developed robust FET models using data-driven techniques, exploring data augmentation strategies including GANs and Autoencoders. Enhanced a limited dataset of 80,000 samples from 27 transistors for improved semiconductor device modeling.

Supervised by Professor Korkut Kaan Tokgoz

Yale University (Remote)

Conducted research on enhancing decision-making by LLMs in multi-agent settings by integrating external tools. Developed the STRIDE framework for strategic and interactive decision-making.

Supervised by Professor Zhuoran Yang and Dr. Chuanhao Li

Sabanci University

Implemented advanced Self-Supervised Learning clustering algorithms (SwAV and BARLOW Twins) to categorize over 7 million Twitter profile pictures, determining distribution of political affiliations and human/bot ratios.

Supervised by Professor Onur Varol