Benchmarking Forensic Reliability: A Comparative Analysis of Automated AI Models versus Human Perception in Detecting Low-Bitrate Deepfakes

DOI

10.2991/978-94-6239-610-4_5How to use a DOI?

Keywords

Deepfake detection; GAN fingerprints; H.264 compression; digital forensics; human–AI comparison

Abstract

The rapid progress of Generative Adversarial Networks (GANs) has enabled highly realistic facial manipulations, creating a serious threat to the integrity of digital evidence. Modern deepfake detectors, particularly convolutional neural network (CNN) based architectures, report strong performance on curated benchmarks; however, their practical reliability is rarely tested under the heavy lossy compression applied by social and messaging platforms. This work presents a comparative forensic study of three automated detection pipelines XceptionNet, MesoNet, and a temporal CNN–RNN model against documented human performance when videos are degraded using H.264 compression at C23 and C40 levels. Using 150 samples from the FaceForensics++ dataset [1], a new metric, the GAN Fingerprint Survivability Index (GFSI), is introduced to quantify the fraction of high-frequency forensic cues that survive compression. Empirically, GFSI ≈ 0.0004 at C40, indicating near-total removal of spectral GAN fingerprints. Experimental results show that XceptionNet collapses into an extreme “fake” bias, MesoNet loses 12% accuracy, and the temporal model performs below random chance on low-quality videos, whereas humans maintain approximately 70% accuracy by relying on semantic cues. These findings demonstrate that automated detectors alone are unsafe for low-bitrate forensic evidence and motivate hybrid human-AI workflows.

Copyright

© 2026 The Author(s)

Open Access

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

Download article (PDF)

TY  - CONF
AU  - Sai Bhavani Venkatesh Pasupuleti
PY  - 2026
DA  - 2026/05/05
TI  - Benchmarking Forensic Reliability: A Comparative Analysis of Automated AI Models versus Human Perception in Detecting Low-Bitrate Deepfakes
BT  - Proceedings of the First International Conference on Advances in Forensics and Cyber Technologies (ICFACT 2025)
PB  - Atlantis Press
SP  - 20
EP  - 26
SN  - 2352-538X
UR  - https://doi.org/10.2991/978-94-6239-610-4_5
DO  - 10.2991/978-94-6239-610-4_5
ID  - Pasupuleti2026
ER  -