Maria Rosaria Briglia

The rapid spread of open-source generative models has made it easy to create highly realistic manipulated media, posing a critical threat to content authenticity and provenance. Proactive image protection methods aim to neutralize this risk, but their assumptions may not hold under modern generative pipelines. We study how diffusion-based upsampling can erase or evade several common protective schemes and quantify the consequences for threat modeling.

Vision–Language Models (VLMs) have become essential for tasks such as image synthesis, captioning, and retrieval by aligning textual and visual information in a shared embedding space. In this work, we show that harmful and benign prompts can be separated more naturally in hyperbolic space, and we exploit that geometry to improve both detection and sanitization.

As modern text-to-image (T2I) models draw closer to synthesizing highly realistic content, the threat of unsafe content generation grows, and it becomes paramount to exercise control over the semantic directions that the model follows. We show that hyperbolic geometry offers a more stable and interpretable basis for semantic steering than standard Euclidean alternatives.

We introduce APEX, a method for anchored protein engineering that combines quantized expectation with structure-aware editing to generate constrained protein variants while preserving desired interface properties.

Accurate modeling of antibody–antigen (Ab–Ag) complexes is central to biologic development, yet the reliability and failures of modern Ab–Ag folding pipelines remain poorly understood. We benchmark large-scale complex folding to identify where models succeed, where they fail, and what this means for engineering and evaluation.

We answer the question in the title, showing that adversarial training (AT) for diffusion models (DMs) fundamentally differs from classifiers: while AT in classifiers enforces output robustness, the challenge in DMs is to preserve data fidelity while controlling the generative trajectory under adversarial perturbations.

CLIP is a discriminative model trained to align images and text in a shared embedding space. Due to its multimodal structure, it serves as the backbone of many generative pipelines, yet the inversion problem remains underexplored. We show that implicit inversion can recover a decoder-like behavior from CLIP, with implications for controllable generation and representation understanding.

We aim at using the Energy-based Model (EBM) framework to better understand adversarial training (AT) in classifiers, and additionally to analyze the intrinsic generative capabilities that emerge from robustly trained discriminative models.

By reinterpreting a robust discriminative classifier as an Energy-based Model (EBM), we offer a new take on the dynamics of adversarial training (AT). Our analysis of the energy landscape provides a clearer explanation of robust overfitting and the geometric behavior of decision boundaries.