- Cross-vector backdoor interference benchmark in QLoRA-fine-tuned LLMs (2026, ongoing). Custom multi-vector poisoning benchmark on Qwen3-4B fine-tuned via QLoRA, with input from Prof. Simone Scardapane. Introduces a new metric (INT_net = ASR_isolated - ASR_concurrent) to quantify whether multiple concurrent backdoors interfere destructively or reinforce each other.
- From Prompts to Perturbations: Adversarial ML at Inference and Training (2024-2025). SSAS research project supervised by Prof. Marco R. Marini and Prof. Emanuele Rodolà. Adversarial threats across vision evasion (FGSM/PGD on ResNet-18), training-time poisoning, backdoor attacks, and prompt injection on tool-augmented OpenAI agents (gpt-4o-mini through gpt-5).
- Geodesic Post-Hoc Quantization for Variational Autoencoders (2025). Independent research project for the Machine Learning course at Sapienza. Post-hoc quantization pipeline for VAEs comparing Euclidean and decoder-induced Riemannian metrics on FashionMNIST and CIFAR-10.
- Employment of Extended Reality and Artificial Intelligence in the Health Care Sector (2023-2024). Qualitative research project with Prof. Emanuele Rodolà and Prof. Marco A. Deriu (Politecnico di Torino), evaluating AI integration, data biases, and scalability in clinical and educational settings.