The presidential deepfakes dataset

Aruna Sankaranarayanan; Matthew Groh; Rosalind Picard; Andrew Lippman

The presidential deepfakes dataset

Aruna Sankaranarayanan^*, Matthew Groh, Rosalind Picard, Andrew Lippman

^*Corresponding author for this work

Management and Organizations

Research output: Contribution to journal › Conference article › peer-review

6 Scopus citations

Abstract

How do we evaluate media forensic techniques for detecting deepfakes? We present the Presidential Deepfakes Dataset (PDD), which consists of 32 videos, half of which are original videos and half of which are manipulated with audio impersonations, synthesized lip synchronizations, political misinformation, and situational artifacts. This dataset expands the context on which end-to-end media forensic systems can be evaluated. As an example, we evaluate the winning model of the DeepFake Detection Challenge on the PDD and find that it classifies 69% of the videos in the PDD accurately. We share this dataset publicly for researchers to evaluate their techniques with the intention of pre-bunking future misinformation attempts.

Original language	English (US)
Pages (from-to)	57-72
Number of pages	16
Journal	CEUR Workshop Proceedings
Volume	2942
State	Published - 2021
Event	1st Workshop on Adverse Impacts and Collateral Effects of Artificial Intelligence Technologies, AIofAI 2021 - Montreal, Canada Duration: Aug 19 2021 → …

Keywords

Dataset
Deepfakes
DFDC
Disinformation
Media forensics
Misinformation
Politics

ASJC Scopus subject areas

General Computer Science

Cite this

@article{6ec71906c95644bebba6d2958a624c97,

title = "The presidential deepfakes dataset",

abstract = "How do we evaluate media forensic techniques for detecting deepfakes? We present the Presidential Deepfakes Dataset (PDD), which consists of 32 videos, half of which are original videos and half of which are manipulated with audio impersonations, synthesized lip synchronizations, political misinformation, and situational artifacts. This dataset expands the context on which end-to-end media forensic systems can be evaluated. As an example, we evaluate the winning model of the DeepFake Detection Challenge on the PDD and find that it classifies 69% of the videos in the PDD accurately. We share this dataset publicly for researchers to evaluate their techniques with the intention of pre-bunking future misinformation attempts.",

keywords = "Dataset, Deepfakes, DFDC, Disinformation, Media forensics, Misinformation, Politics",

author = "Aruna Sankaranarayanan and Matthew Groh and Rosalind Picard and Andrew Lippman",

note = "Publisher Copyright: {\textcopyright} 2021 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).; 1st Workshop on Adverse Impacts and Collateral Effects of Artificial Intelligence Technologies, AIofAI 2021 ; Conference date: 19-08-2021",

year = "2021",

language = "English (US)",

volume = "2942",

pages = "57--72",

journal = "CEUR Workshop Proceedings",

issn = "1613-0073",

publisher = "CEUR-WS",

}

TY - JOUR

T1 - The presidential deepfakes dataset

AU - Sankaranarayanan, Aruna

AU - Groh, Matthew

AU - Picard, Rosalind

AU - Lippman, Andrew

PY - 2021

Y1 - 2021

N2 - How do we evaluate media forensic techniques for detecting deepfakes? We present the Presidential Deepfakes Dataset (PDD), which consists of 32 videos, half of which are original videos and half of which are manipulated with audio impersonations, synthesized lip synchronizations, political misinformation, and situational artifacts. This dataset expands the context on which end-to-end media forensic systems can be evaluated. As an example, we evaluate the winning model of the DeepFake Detection Challenge on the PDD and find that it classifies 69% of the videos in the PDD accurately. We share this dataset publicly for researchers to evaluate their techniques with the intention of pre-bunking future misinformation attempts.

AB - How do we evaluate media forensic techniques for detecting deepfakes? We present the Presidential Deepfakes Dataset (PDD), which consists of 32 videos, half of which are original videos and half of which are manipulated with audio impersonations, synthesized lip synchronizations, political misinformation, and situational artifacts. This dataset expands the context on which end-to-end media forensic systems can be evaluated. As an example, we evaluate the winning model of the DeepFake Detection Challenge on the PDD and find that it classifies 69% of the videos in the PDD accurately. We share this dataset publicly for researchers to evaluate their techniques with the intention of pre-bunking future misinformation attempts.

KW - Dataset

KW - Deepfakes

KW - DFDC

KW - Disinformation

KW - Media forensics

KW - Misinformation

KW - Politics

UR - http://www.scopus.com/inward/record.url?scp=85114698020&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85114698020&partnerID=8YFLogxK

M3 - Conference article

AN - SCOPUS:85114698020

SN - 1613-0073

VL - 2942

SP - 57

EP - 72

JO - CEUR Workshop Proceedings

JF - CEUR Workshop Proceedings

T2 - 1st Workshop on Adverse Impacts and Collateral Effects of Artificial Intelligence Technologies, AIofAI 2021

Y2 - 19 August 2021

ER -

The presidential deepfakes dataset

Abstract

Keywords

ASJC Scopus subject areas

Other files and links

Fingerprint

Cite this