FakeParts: a New Family of AI-Generated DeepFakes

TL;DR

This work identifies a worrying gap in deepfake sensing: localized, partial manipulations within otherwise authentic videos (FakeParts). It defines FakeParts, introduces FakePartsBench—the first large-scale benchmark incorporating full deepfakes and diverse FakeParts with fine-grained annotations—and provides extensive human and detector evaluations. Findings show detectors, especially out-of-the-box and language-model–based ones, struggle with partial edits, while diffusion-based detectors offer the most robust but still imperfect performance. The results underscore a pressing need for localization-aware defenses and detailed benchmarks to advance media integrity in the era of diffusion-driven video generation.

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.

Figures (15)

• Figure 1: FakePartsBench is the first dataset specifically designed to include both full deepfakes and FakeParts.
• Figure 2: Prompt word-count.
• Figure 3: Pipeline of FakePartsBench includes both full and FakeParts Deepfake videos. The partial manipulations, FakeParts, are categorized into three types: temporal, spatial, and style.
• Figure 4: FakeParts examples grouped by manipulation group (per column) with three samples per group.
• Figure 5: Results in Radar map of FakePartsBench performance across detector categories, where for each category we report the average performance achieved among all generators: OoB, FoM, VLM, and DiB and on various generators, including full fake generators such as: T2V, TI2V, and FakeParts: Change-of-Style, Inpainting, Outpainting, Interpolation, Extrapolation, and Faceswap.