CoChat

Multisemantic Path Neural Network for Deepfake Detection

Open access
Wu, Nan, Jin, Xin, Jiang, Qian, Wang, Puming, Zhang, Ya, Yao, Shaowen, Zhou, Wei · 2022 · Security and Communication Networks

The DOI matches its Crossref record; 61 of the 61 references it cites were found in a registry or an index; 1 of the 15 sources its claims cite were checked for support.

Abstract

With the continuous development of deep learning techniques, it is now easy for anyone to swap faces in videos. Researchers find that the abuse of these techniques threatens cyberspace security; thus, face forgery detection is a popular research topic. However, current detection methods do not fully use the semantic features of deepfake videos. Most previous work has only divided the semantic features, the importance of which may be unequal, by experimental experience. To solve this problem, we propose a new framework, which is the multisemantic pathway network (MSPNN) for fake face detection. This method comprehensively captures forged information from the dimensions of microscopic, mesoscopic, and macroscopic features. These three kinds of semantic information are given learnable weights. The artifacts of deepfake images are more difficult to observe in a compressed video. Therefore, preprocessing is proposed to detect low-quality deepfake videos, including multiscale detail enhancement and channel information screening based on the compression principle. Center loss and cross-entropy loss are combined to further reduce intraclass spacing. Experimental results show that MSPNN is superior to contrast methods, especially low-quality deepfake video detection.

Overview

CoChat Summary

TL;DR

The paper proposes MSPNN, a multisemantic path neural network that fuses microscopic, mesoscopic, and macroscopic features with learnable channel-spatial attention weights for deepfake detection. It reports strong accuracy on FaceForensics++ and DeepFake-TIMIT, especially for low-quality/high-compression videos, with cross-dataset generalization on Celeb-DFv2.

Key findings

  • MSPNN achieved 94.21% accuracy on FF++(HQ/C23) and 76.31% on FF++(LQ/C40), and AUCs of 99.12 (TIMIT HQ) and 99.52 (TIMIT LQ), outperforming the compared methods in Table 5.
  • Cross-dataset validation on Celeb-DFv2 gave an AUC of 66.7, while the same model scored 96.7 on FF-DF; the paper notes Masi’s method generalizes better on Celeb-DFv2 (73.41) but has a much lower original-dataset AUC.
  • Ablations show preprocessing and attention help some subsets, especially high-compression data: Y-channel input raised FF++(LQ) NeuralTextures from 74.41% (RGB) to 74.92%, and attention raised FF++(LQ) DeepFakes from 93.83% to 94.87% (Tables 2 and 4).
  • The three semantic pathways are intended to capture different forgery cues: microscopic details, mesoscopic blocks, and macroscopic regions such as eyes, nose, and mouth, with weights assigned automatically rather than manually.

Methods

  • Design: MSPNN uses three semantic pathways based on receptive fields—micro (3×3), meso (two 3×3 replacing 5×5), and macro (three 3×3 replacing 7×7)—plus multiscale detail enhancement with three Gaussian filters (5×5, 9×9, 19×19; σ = 1.0, 2.0, 4.0; weights 0.5, 0.5, 0.25) and YCrCb channel separation using the Y channel. Features are fused via channel-spatial attention, and training combines cross-entropy loss with center loss (λ = 1).
  • Data: FaceForensics++ (1,000 YouTube videos; 1,000 fake and 1,000 real videos per compression rate; splits 720 train/140 validation/140 test), DeepFake-TIMIT (640 fake videos at 128×128 and 64×64; splits 230/45/45), and Celeb-DFv2 (590 real, 5,639 deepfake videos) for cross-dataset validation. Faces were extracted with dlib, expanded by 1.3×, and 30 frames per video were taken.
  • Implementation/analysis: PyTorch on an RTX 3090; Adam with initial learning rate 1e-4 and cosine annealing, center loss optimized with SGD, batch size 32, 100 epochs, input 224×224, random seed 43. Evaluation used accuracy and AUC, with preprocessing, pathway, attention, and cross-dataset ablations.

Limitations

  • The paper states that balancing accuracy and generalization remains a challenge; Masi’s method generalizes better on Celeb-DFv2 than MSPNN, though its original-dataset AUC is far behind.
  • The paper states that only intraframe compression was used because removing adjacent frames would conflict with benchmark dataset creation requirements; dlib sometimes failed to recognize faces, and those frames were skipped.
  • Evident from the tables: gains from detail enhancement and attention are mixed, with some subsets decreasing (e.g., detail enhancement lowered FF++ HQ DeepFakes from 99.73% to 99.54%, and attention lowered FF++ HQ DeepFakes from 99.73% to 99.49%), so improvements are not uniform across all datasets or compression levels.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1155/2022/4976848

Cites

61 of 61 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

1 Valid · 3 Not valid · 11 Not checked · 46 Own results and interpretations

25 more claims

Create a free account to see all 61 claims and to verify the 14 references still open.

See all claims — free account
  • It may state two separate claims as one. 19
  • Probably not a substantial claim. 13

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

4 of 4

Citations

Cites

Show all 61 references

Research works printed but not linked: 2

  1. Y. Jo, J. Park, and Sc-fegan, "Face editing generative adver- sarial network with user's sketch and color," in Proceedings of the 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 1745-1753, Seoul, Korea, November 2019.
  2. L. Li, C. Chen, L. Pan, J. Zhang, and Y. Xiang, "Sok: an overview of ppg's application in authentication," 2022, [Online]. Available: https://arxiv.org/abs/2201.11291. · arXiv

Cited by

Metadata

Authors
Wu, N., Jin, X., Jiang, Q., Wang, P., Zhang, Y., Yao, S., & Zhou, W.
Affiliations
Yunnan University (CN)
Corresponding author
Jiang, Qian
Venue
Security and Communication Networks
Publisher
Wiley
Published
2022-10-11
Volume
2022
Pages
1-14
ISSN
  • 1939-0114 (ISSN-L)
  • 1939-0122
OpenAlex
W4304587257
Cited by
9
References
55
Citations by year
2026: 1 · 2025: 5 · 2024: 1 · 2023: 2
FWCI
0.65
Type
journal article
Language
en
Open access
Locations
  • Security and Communication Networks (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
Funders
National Natural Science Foundation of China (202202AD080002, 202201AT070112, 62166047, 62002313, 62002313, 62101481, 62101481, 2020SE408, 2021–2022)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Generative Adversarial Networks and Image Synthesis
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Image and Video Quality Assessment
Keywords
deepfake detection, face forgery detection, semantic features, video compression artifacts, center loss, face swapping

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-07, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-07 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.