Salvatore Sambataro

I am a PhD student in Artificial Intelligence and Computer Vision, focusing on Generative AI, Video Understanding, and Multimedia Forensics. My research explores how to analyze and interpret complex visual content from videos, as well as the capabilities, limitations, and societal impact of modern generative models and synthetic media. I design methods for video understanding that combine deep learning, representation learning, and multimodal reasoning to extract meaningful information from dynamic visual data.

Alongside my academic work, I am an AI Engineer and Researcher at iCTLab s.r.l., where I develop and deploy AI solutions across a range of domains. I primarily work on computer vision projects, but I also contribute to systems involving time series analysis, natural language processing, and retrieval-augmented generation (RAG). In this role, I bridge research and production: I design models, build scalable pipelines, and translate cutting-edge ideas into real-world applications. I am also actively involved in research activities within the company, contributing to experimental prototypes, publications, and collaborative innovation.

news

Sep 25, 2026	Our paper SynthForensics: Benchmarking and Evaluating People-Centric Synthetic Video Deepfakes has been accepted as a poster at the NeurIPS 2026 Evaluations & Datasets Track! [arXiv]
Sep 08, 2026	Our challenge paper Adversarial Attacks on Deepfake Detectors: A Challenge in the Era of AI-Generated Media (AADD-2026) has been accepted at ACM Multimedia 2026 in Rio de Janeiro! [Challenge]
May 08, 2026	Major update of SynthForensics on arXiv: now 20,445 videos from 8 T2V and 7 I2V generators, a human preference study and 15 evaluated detectors.
Feb 04, 2026	The first version of SynthForensics, a human-centric benchmark for detecting synthetic video deepfakes, is now available on arXiv!

selected publications

NeurIPS

SynthForensics: Benchmarking and Evaluating People-Centric Synthetic Video Deepfakes

Roberto Leotta^† Salvatore Alfio Sambataro^†, and 6 more authors

In Advances in Neural Information Processing Systems (NeurIPS), Evaluations & Datasets Track, 2026

⭐ Poster · † Joint first authorship

Abs arXiv

Modern T2V/I2V generators synthesize people increasingly hard to distinguish from authentic footage, while current evaluation suites lag: legacy benchmarks target manipulation-based forgeries, and recent synthetic-video benchmarks prioritize scale over realistic human depiction. We introduce SynthForensics, a people-centric benchmark of 20,445 videos from 8 T2V and 7 I2V open-source generators, paired-source from FF++/DFD reals, two-stage human-validated, in four compression versions with full metadata. In our paired-comparison human study, raters prefer SynthForensics in 71–77% of head-to-head comparisons against each of nine existing synthetic-video benchmarks, while facial-quality metrics fall within the FF++/DFD baseline range. Across 15 detectors and three protocols, face-based methods drop 13–55 AUC points (mean 27) from FF++ to SynthForensics and a further 23 under aggressive compression; fine-tuning closes the gap at a backward cost on legacy benchmarks; training from scratch shows synthetic and manipulation features largely disjoint for most detectors. We release dataset, pipeline, and code.
ACM MM

Adversarial Attacks on Deepfake Detectors: A Challenge in the Era of AI-Generated Media

Francesco Guarnera, Luca Guarnera, Andrea Montibeller, Salvatore Alfio Sambataro, and 6 more authors

In Proceedings of the 34th ACM International Conference on Multimedia (MM ’26), 2026

Abs DOI PDF Website

The increasing realism of AI-generated media and its widespread dissemination through social media demand deepfake detectors that remain reliable under adversarial attacks and realistic post-processing. We present AADD-2026, a competitive benchmark for evaluating adversarial attacks against deepfake detectors under realistic deployment conditions. The challenge considers synthetic images from five generative families, social-media-inspired processing, JPEG and JPEG-AI compression, and eight heterogeneous detectors operating across spatial, frequency, and foundation-model representations. Ten teams submitted valid solutions, exploring latent-space, frequency-aware, ensemble, manifold-projection, and compression-aware attacks. Results show that explicitly accounting for compression and post-processing substantially improves transferability to unseen detectors. The winning approach achieves the highest attack success through a gradient-free real-face manifold projection strategy. Meanwhile, DINOv2- and SigLIP2-based detectors show substantially greater resilience than conventional architectures, highlighting a clear trade-off between attack transferability and perceptual fidelity.