Research Scientist, Content Authenticity Initiative
Adobe
I’m a research scientist in the AI for Content Authenticity team at Adobe. My primary research interest lies in content provenance and multimedia forensics, with a focus on both active and passive forensic techniques such as watermarking, deepfake detection, and robust image fingerprinting.
Previously, I was a postdoc at MIT CSAIL with Prof. William Freeman, and at the same time a Lecturer for the Computer Vision course in the MIDS program at the UC Berkeley School of Information. I did my PhD in Computer Science at UC Berkeley, advised by Prof. Hany Farid.
I organize workshops and tutorials on media forensics and watermarking (CVPR, ICCV, ACM Multimedia, IH&MMSec), and my work on TrustMark was recognized with an Adobe Tech Excellence Award.
Content provenance is information on the creation, authorship, and editing of a digital asset. Content Credentials carry this as metadata with cryptographic means to verify it is unchanged. But what happens when platforms strip the metadata? Durable Content Credentials unify three pillars — metadata, invisible watermarking, and image fingerprinting — so that credentials stay permanently attached and are safeguarded against spoofing.
Sophisticated fake videos were once the domain of Hollywood studios or state actors; deep learning has democratized them. Easy access to technology that can make anybody say anything threatens elections, fuels fraud and misinformation, and enables non-consensual pornography. My work develops forensic techniques to detect such synthetic media.
Facial images serve as proof of identity from social media to national ID documents, and synthesis and manipulation make fake identities increasingly easy to create. Here we run perceptual studies to measure how well humans distinguish original faces from synthetic and manipulated ones.
Directed rounding of DCT coefficients in JPEG encoders leaves a periodic artifact — a darker or brighter pixel in each 8×8 block. Analysis of thousands of camera models shows it is widespread, and its local absence reveals manipulations such as content-aware fill, resampling and compositing.
Camera-based forensics can fail when a manipulated image is re-imaged (photographing a print or a display, scanning, or screen-grabbing). We collected a large-scale dataset of all four attack types and trained a CNN that substantially outperforms prior techniques.
Surveillance license-plate images are often noisy and tiny, around 20 pixels wide. Large-scale simulation showed useful information remains even at 1.9 pixels per character, and a trained CNN recognizes characters on degraded real-world plates.
S. Agarwal, T. Bui, J. Collomosse
S. Agarwal, E. Michel, V. Asnani, T. Mathern, J. Collomosse
L. Zhang, S. Agarwal, J. Collomosse, P. Xie, V. Asnani
V. Asnani, J. Collomosse, T. Bui, X. Liu, S. Agarwal
S. Agarwal, L. Hu, E. Ng, T. Darrell, H. Li, A. Rohrbach
V. Asnani, J. Collomosse, X. Liu, S. Agarwal