FakeParts: a New Family of AI-Generated DeepFakes

View on arXiv ← Back to list

Authors: Ziyi Liu, Firas Gabetni, Awais Hussain Sani, Xi Wang, Soobash Daiboo, Gaetan Brison, Gianni Franchi, Vicky Kalogeiton

Published: 2025-08-28 17:55:14+00:00

AI Summary

This paper introduces FakeParts, a new class of deepfakes characterized by subtle, localized manipulations to specific spatial regions or temporal segments of otherwise authentic videos. To address this challenge, the authors present FakePartsBench, the first large-scale benchmark dataset specifically designed for partial deepfakes, featuring over 81K videos with pixel- and frame-level annotations. Their studies reveal that FakeParts significantly reduce both human and state-of-the-art deepfake detector accuracy, highlighting an urgent vulnerability in current detection methods.

Abstract

We introduce FakeParts, a new class of deepfakes characterized by subtle, localized manipulations to specific spatial regions or temporal segments of otherwise authentic videos. Unlike fully synthetic content, these partial manipulations - ranging from altered facial expressions to object substitutions and background modifications - blend seamlessly with real elements, making them particularly deceptive and difficult to detect. To address the critical gap in detection, we present FakePartsBench, the first large-scale benchmark specifically designed to capture the full spectrum of partial deepfakes. Comprising over 81K (including 44K FakeParts) videos with pixel- and frame-level manipulation annotations, our dataset enables comprehensive evaluation of detection methods. Our user studies demonstrate that FakeParts reduces human detection accuracy by up to 26% compared to traditional deepfakes, with similar performance degradation observed in state-of-the-art detection models. This work identifies an urgent vulnerability in current detectors and provides the necessary resources to develop methods robust to partial manipulations.

Key findings

FakeParts significantly reduce human detection accuracy by up to 26% and cause similar performance degradation in state-of-the-art deepfake detectors, identifying a critical vulnerability. While full deepfakes are comparatively easy for most detectors, existing Out-of-the-Box, VLM-based, and Foundation-Model-based methods generally perform poorly or are brittle on FakeParts. Diffusion-based detectors offer the most balanced performance across full and partial deepfakes, though challenges remain, particularly with temporal manipulations like interpolation and extrapolation.

Approach

The authors define and characterize FakeParts as a novel family of deepfakes with subtle, localized spatial or temporal video manipulations. To enable robust detection research, they create and release FakePartsBench, a large-scale benchmark dataset comprising both full deepfakes and various types of FakeParts with fine-grained pixel- and frame-level annotations. They then conduct comprehensive human and detection studies using this benchmark to evaluate the performance of state-of-the-art deepfake detectors against these new, more deceptive manipulations.

Datasets

FakePartsBench, DAVIS 2016, DAVIS 2017, YouTube-VOS 2019, MOSE, LVD-2M, Celeb-DF, CelebA, Animal Kingdom, Pexels, WebVid-10M

Model(s)

UNKNOWN

Author countries

France

← Previous