Istanbul, Türkiye

Mert Coşkuner

MSc student in Computer Science & Engineering, Sabancı University

I work on trustworthy and efficient machine learning: how learning systems behave under attack and under constraints. My recent projects cover Byzantine-robust and differentially private federated learning, backdoor attacks on graph neural networks, adversarial robustness and knowledge distillation, and 4-bit quantization of multilingual language models.

I am currently a researcher at TÜBİTAK BİLGEM and a teaching assistant for Security, Privacy and Robustness in Machine Learning at Sabancı University.

GitHub Google Scholar LinkedIn Email
Research

Interests

  • Federated learning under Byzantine attacks and differential privacy
  • Backdoor and evasion attacks, and defenses, on vision and graph models
  • Robustness of compressed models: distillation and low-bit quantization
  • Multilingual and Turkish language models
Selected publications

All publications

Selected work
Figure from Structure-aware FedALA
Federated graph learning 2026

Structure-aware FedALA

Personalised federated graph learning with topology-weighted aggregation, a proximal term and degree-calibrated differential privacy, benchmarked on seven graph datasets with 10 and 20 clients.

federated learningGNNdifferential privacy
Figure from TR-QuantBench
Efficient LLMs 2026

TR-QuantBench

How the language of the calibration data shapes 4-bit GPTQ and AWQ quantization of multilingual LLMs, with a Turkish focus and a tokenizer-density analysis.

LLMsquantizationTurkish NLP
Figure from Backdoors in graph neural networks
ML security 2026

Backdoors in graph neural networks

A PyTorch Geometric framework for graph, node and link tasks with Graph Trojaning Attack backdoors on graph classification, plus spectral, topological and feature analyses of poisoned graphs.

GNNbackdoor attacksPyG
Figure from CNN robustness on CIFAR-10
Adversarial robustness 2026

CNN robustness on CIFAR-10

Corruption and adversarial robustness of transfer-learned and distilled CNNs on CIFAR-10/-C: AugMix with JSD, PGD, Grad-CAM, and how a teacher's adversarial examples transfer to its students.

robustnessdistillationCIFAR-10-C
Figure from Transformer feedback channel coding
Learning for communication 2026

Transformer feedback channel coding

End-to-end learned channel codes for the AWGN channel with noiseless feedback: a transformer transmitter and receiver trained through the channel, with ablations over rounds, noise and capacity.

transformerschannel coding
Figure from Aspect-based recommender
NLP 2026

Aspect-based recommender

Product recommendation from Amazon reviews via instruction-tuned aspect-based sentiment analysis, LDA and BERTopic topics, and hierarchical meta-aspect clustering.

ABSAtopic modelsclustering

All projects

Experience
Teaching
Education