Unmasking Neural Puppetry: Detection of Lip-Syncing Deepfakes \in-the-wild" with Meaningful Explanations
Keywords
Loading...
Authors
Issue Date
2021-06-01
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
We have entered an era of misinformation and fake news, fuelled by Arti cial Intelligence
(AI). In the visual media, malicious entities have weaponized synthetic multimedia using
the Deepfake technology. Within deepfake technology, state-of-the-art (SOTA) lipsyncing
deepfake techniques have evolved to \out-of-the-box" target speaker-agnostic
manipulations, i.e. they do not require long hours of training videos of a target speaker.
Given the societal impact of such commodi cation of creating lip-syncing deepfakes, our
research proposes learning-based methods to detect lip-syncing deepfake videos found
\in-the-wild". With no prior work into target speaker-agnostic lip-sync deepfake detection,
we introduce a new synthetic dataset for our research and select the E cientNet-B4
architecture, SOTA in deepfake detection research, for the lip-sync deepfake detection
task. The baseline detector model achieves accuracy = 0:78 and FPR = 0:12 on our
curated evaluation benchmark, a collection of lip-syncing deepfake videos found \in-thewild".
Within the context of improving detector model performance on the evaluation
benchmark, we demonstrate transfer learning from a SOTA face-swap deepfake detector
model; and achieve our best performing lip-sync deepfake detector model using Center
loss, with accuracy = 0:88 and FPR = 0:09 on our evaluation benchmark.
With limited research in explainable deepfake detection, we contribute towards both
extrinsic and intrinsic explainability in the lip-sync deepfake detection task. We use a
perturbation-based input attribution method, the Occlusion algorithm, to provide posthoc
model explanations to the end user and regulatory stakeholder groups. We also
demonstrate that such post-hoc model explanations can be used a feedback mechanism
to debug model behaviour; and draw insights for improvements in model performance
on the evaluation benchmark. Finally, we propose new research into intrinsic explainability
in deepfake detection using Concept Whitening. With no prior work into de ning
\concepts" for deepfake manipulations, we use Concept Whitening as an unsupervised
learning strategy on the E cientNet-B4 architecture to identify lip-sync deepfake manipulations
as \concepts". We achieve concept-based model interpretability on our detector
model without compromising performance on the evaluation benchmark, with
accuracy = 0:86 and FPR = 0:06; and demonstrate that the learned \concepts" can
be presented as meaningful explanations when visualised as saliency maps. Hence, we
establish Concept Whitening as a viable technique for further research into interpretable
deepfake detection.
Keywords - media forensics, deepfake detection, lip-sync, neural network, explainable
AI
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen