🎓 BSc in Applied Computer Science and Data Analytics — University of Cagliari, July 2026
🎯 MSc Computer Engineering, Cybersecurity & AI (CyberAI) — UniCa, from September 2026
📍 Cagliari, Sardinia, Italy | 🌍 Open to remote and international opportunities
📧 davide.deplano@gmail.com
🔗 LinkedIn
Technosignature Search at INAF – Sardinia Radio Telescope (9 months)
Adaptation and validation on real SRT data of a multi-stage anomaly-detection pipeline for SETI-like radio signals, based on the methodology of Pardo et al. (AJ 2025). Candidates pass a density-based pre-filter (cross-correlation features, UMAP embedding, per-category KDEs) trained on simulated cadences, are then scored by a bagged-GMM frequency filter and an ON/OFF similarity metric in UMAP space, and ranked to prioritise human inspection.
Main contribution: an NCC-max cross-correlation extractor tolerant to frequency drifts, validated against a Pearson baseline. Two exploratory branches — a Keras Tuner–optimised convolutional autoencoder and a Random Forest on handcrafted ON/OFF contrast features — were also implemented; the Random Forest result is a documented negative one, high accuracy on the synthetic domain and indistinguishable score distributions on real data, which quantifies the simulation-to-reality gap of the training set.
Built as a reproducible framework: YAML configuration, pluggable filter interface, UML documentation, memory-mapped training for 1.28M synthetic cadences, and an interactive UMAP reverse-search tool for cluster and RFI investigation.
Synthetic speech detector: frozen wav2vec2 encoder with a trained classification head, served as a containerised FastAPI service with tests and CI.
The project measures the generalisation gap rather than chasing accuracy. The same model scores 3.66% EER on known attacks in studio conditions, 8.52% on thirteen unseen attack algorithms, and 29.66% on real-world audio. Quadrupling the training data cut in-domain error by 29% and out-of-domain error by 5.7% — the gap widened rather than closing, showing the failure is domain shift, not undertraining.
Binary CNN deepfake detector (RGB vs grayscale) evaluated under PGD white-box and transfer black-box attacks, hardened with PGD adversarial training (Madry et al., 2018).
| Model | Clean | PGD ε=0.01 | PGD ε=0.02 |
|---|---|---|---|
| RGB, undefended | 97.9% | 62.1% | 9.6% |
| RGB, adversarially trained | 99.2% | 87.9% | 41.3% |
| Grayscale, undefended | 98.3% | 71.7% | 20.8% |
| Grayscale, adversarially trained | 97.9% | 85.4% | 47.1% |
Imperceptible perturbations collapse undefended detectors from ~98% to ~62–72%; adversarial training recovers ~26 points on RGB at ε=0.01 at negligible cost in clean accuracy. At ε=0.05 the defense no longer generalises — both models still collapse.
Network intrusion detection on KDD Cup 1999 (10% subset), comparing supervised and unsupervised approaches with hyperparameter tuning and feature interpretation.
Random Forest F1 ≈ 0.999 · One-Class SVM F1 ≈ 0.994 · MLP F1 ≈ 0.990 · VAE F1 ≈ 0.969
Credit card payment default prediction on the UCI Default of Credit Card Clients dataset (22% positive class). Random Forest, KNN, Naive Bayes and two custom classifiers, each evaluated across five preprocessing configurations (scaling, variance-threshold selection, SMOTE balancing), with GridSearchCV tuning optimised for F1 on the minority class.
Random Forest is the strongest model: accuracy 0.82, ROC-AUC 0.76. SMOTE raises recall on the default class from 0.37 to 0.47 (F1 0.50) at the cost of precision — the recall/precision trade-off that matters in credit risk.
Software-engineering project (with Sebastiano Seu): a composable data pipeline over simulated aquarium sensor data, computing per-tank stress metrics and statistics on large volumes.
Built as production-style Python rather than a script: uv with a lockfile for reproducible
installs, unit and acceptance test suites under pytest, static typing checked with mypy,
linting with ruff, pre-commit hooks, and a GitHub Actions workflow running lint, type-check and
tests on every push and pull request. Optional run logging to Weights & Biases. Designed from UML
use-case, class and sequence diagrams.
Performance work with the Scalene profiler: the hot path pairwise_stress_function is
JIT-compiled with Numba and parallelised with Joblib, bringing computations.py down to 14.3% of
total runtime (966 ms of 6.765 s); the interactive profiling report is committed to the repo.
Data handling uses Polars.
Machine Learning & Deep Learning — CNNs, autoencoders, VAEs, GMMs, manifold learning (UMAP), tree-based and kernel methods. Supervised and unsupervised pipelines on real data.
AI Security — Adversarial robustness of deep models: evasion attacks (PGD, transfer) and adversarial training defenses.
Data Science — End-to-end pipelines: preprocessing, feature extraction, anomaly detection, evaluation on imbalanced data.
Software Engineering — Modular Python packages, configuration-driven design, unit and acceptance testing with pytest, static typing (mypy), linting (ruff), CI with GitHub Actions, REST services with FastAPI, containerisation with Docker, UML design documentation, and profiling-driven optimisation.
Python TensorFlow/Keras scikit-learn imbalanced-learn UMAP pandas NumPy Git SQL C
⭐ Feel free to explore my pinned repositories or connect on LinkedIn.