Assoc. Prof. Dr. Clara Grazian | Statistics | Best Researcher Award

1. Assessing the Invertibility of Deep Biometric Representations: Investigating CNN Hyperparameters for Enhanced Security Against Adversarial Attacks

Authors: C. Grazian, Q. Jin, G. Tangari
Published in: Expert Systems with Applications, Volume 264, 2025, Article 125848
Summary:
This paper investigates the security vulnerabilities in deep biometric systems by evaluating the invertibility of biometric feature representations derived from Convolutional Neural Networks (CNNs). The authors systematically analyze how different CNN hyperparameters affect the robustness of these models against adversarial inversion attacks. The work proposes tuning strategies to improve security without compromising biometric performance.
Contribution: Enhances understanding of CNN-based biometric security, a crucial area in identity verification systems.
Relevance: AI security, adversarial robustness, biometrics.

2. Darwin 1.5: Large Language Models as Materials Science Adapted Learners

Authors: T. Xie, Y. Wan, Y. Liu, Y. Zeng, S. Wang, W. Zhang, C. Grazian, C. Kit, et al.
Published in: arXiv preprint arXiv:2412.11970, 2024
Summary:
This work introduces Darwin 1.5, a tailored version of large language models (LLMs) specifically adapted for materials science learning tasks. The model is fine-tuned on scientific texts and datasets related to materials discovery, showcasing improvements in knowledge retrieval, data interpretation, and hypothesis generation.
Contribution: Dr. Grazian contributed Bayesian modeling insights to the model evaluation metrics.
Relevance: Interdisciplinary AI application, materials informatics, LLM adaptation.

3. Approximate Bayesian Computation with Statistical Distances for Model Selection

Authors: C. Angelopoulos, C. Grazian
Published in: arXiv preprint arXiv:2410.21603, 2024
Summary:
The paper explores model selection under Approximate Bayesian Computation (ABC) by incorporating robust statistical distance measures (e.g., Wasserstein, Energy Distance). The approach helps mitigate issues in likelihood-free inference where traditional ABC may struggle with model choice accuracy.
Contribution: Dr. Grazian co-developed the methodological framework and designed experiments for evaluating model selection efficacy.
Relevance: Computational statistics, Bayesian inference, ABC methods.

4. Parametric Maps of Kinetic Heterogeneity and Ki in Dynamic Total Body PET using Approximate Bayesian Computation

Authors: Q. Gu, G. Angelis, D. Bailey, P. Roach, C. Grazian, G. Emvalomenos, et al.
Presented at: 2024 IEEE Nuclear Science Symposium (NSS) and Medical Imaging Conference (MIC)
Summary:
This paper applies ABC methods to generate parametric maps from dynamic total-body PET scans, providing estimates for kinetic heterogeneity and Ki (influx rate constant). The approach addresses complex likelihoods in dynamic PET data.
Contribution: Dr. Grazian contributed the statistical modeling and implementation of the ABC framework.
Relevance: Medical imaging, Bayesian computation, PET quantification.

5. Novel Bayesian Algorithms for ARFIMA Long-Memory Processes: A Comparison Between MCMC and ABC Approaches

Authors: J.C. Gabor, C. Grazian
Published in: arXiv preprint arXiv:2410.13261, 2024
Summary:
The study compares traditional MCMC techniques and ABC for estimating parameters of ARFIMA (Autoregressive Fractionally Integrated Moving Average) processes, which model long-range dependencies in time series. The paper highlights the efficiency and trade-offs of both approaches in complex likelihood environments.
Contribution: Dr. Grazian led the design of the ABC-based inference strategy and performance benchmarking.
Relevance: Time series analysis, long-memory processes, Bayesian methodology.

.Conclusion: