DeepFake Detection with Inconsistent Head Poses: Reproducibility and Analysis

TL;DR

This paper critically examines a popular DeepFake detection method based on head pose estimation, revealing significant overstatements of its effectiveness and providing insights into facial landmark detection and algorithmic bias.

Contribution

The study reproduces and analyzes an existing head pose-based DeepFake detector, uncovering overestimated performance claims and offering new insights into related technical challenges.

Findings

The method's effectiveness is significantly overstated in current literature.

Facial landmark detection and head pose estimation are subject to biases affecting detection accuracy.

Reproducibility issues highlight the need for more robust DeepFake detection techniques.

Abstract

Applications of deep learning to synthetic media generation allow the creation of convincing forgeries, called DeepFakes, with limited technical expertise. DeepFake detection is an increasingly active research area. In this paper, we analyze an existing DeepFake detection technique based on head pose estimation, which can be applied when fake images are generated with an autoencoder-based face swap. Existing literature suggests that this method is an effective DeepFake detector, and its motivating principles are attractively simple. With an eye towards using these principles to develop new DeepFake detectors, we conduct a reproducibility study of the existing method. We conclude that its merits are dramatically overstated, despite its celebrated status. By investigating this discrepancy we uncover a number of important and generalizable insights related to facial landmark detection, identity-agnostic head pose estimation, and algorithmic bias in DeepFake detectors. Our results correct the current literature's perception of state of the art performance for DeepFake detection.