Student Publications [Scholarly]

The Generalization Crisis and the Forensic Triad: A Comprehensive Survey of Deepfake Detection (2023–2025)

Document Type

Conference Proceeding

Abstract

The rapid transition from GAN-based to diffusion-based generative AI has precipitated a “Generalization Crisis” in deepfake forensics. This survey provides a systematic review of 40 seminal papers published between 2023 and 2025 to diagnose why state-of-the-art detectors are failing. Our quantitative analysis reveals a catastrophic performance gap: while Phase I artifact-centric models (e.g., Xception, MesoNet) achieve near-perfect accuracy ∼99.2% on legacy benchmarks, they suffer a collapse to random performance (∼50% AUC) when evaluated on contemporary “in-the-wild” datasets. To address this, we propose a new taxonomy categorizing detection methodologies into three evolutionary phases: Artifact-Centric (obsolete GAN-fingerprinting), Foundation-Model Adapted (leveraging frozen backbones like CLIP to restore robustness to 91%), and Semantic-Physical (detecting reality violations). Furthermore, we argue that binary classification is no longer sufficient for forensic utility. We identify the field’s necessary pivot toward a “Forensic Triad”: Attribution, Localization, and Explanation, and outline the critical research gaps preventing current foundation models from meeting these evidentiary standards. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2027.

Publication Title

Lecture Notes in Networks and Systems

Publication Date

2027

Volume

2108 LNNS

First Page

408

Last Page

422

ISBN

9783032341310

DOI

10.1007/978-3-032-34132-7_36

Keywords

adversarial robustness, deepfake detection, foundation models, generalization crisis, Generative AI forensics, multimodal learning

Link to Full Text

Share

COinS