CoChat

FL-TENB4: A Federated-Learning-Enhanced Tiny EfficientNetB4-Lite Approach for Deepfake Detection in CCTV Environments

Open access
Ha, Jimin, El Azzaoui, Abir, Park, Jong Hyuk · 2025 · Sensors

The DOI matches its Crossref record; 29 of the 29 references it cites were found in a registry or an index; 6 of the 14 sources its claims cite were checked for support.

Abstract

The widespread deployment of CCTV systems has significantly enhanced surveillance and public safety across various environments. However, the emergence of deepfake technology poses serious challenges by enabling malicious manipulation of video footage, compromising the reliability of CCTV systems for evidence collection and privacy protection. Existing deepfake detection solutions often suffer from high computational overhead and are unsuitable for real-time deployment on resource-constrained CCTV cameras. This paper proposes FL-TENB4, a Federated-Learning-enhanced Tiny EfficientNetB4-Lite framework for deepfake detection in CCTV environments. The proposed architecture integrates Tiny Machine Learning (TinyML) techniques with EfficientNetB4-Lite, a lightweight convolutional neural network optimized for edge devices, and employs a Federated Learning (FL) approach for collaborative model updates. The TinyML-based local model ensures real-time deepfake detection with minimal latency, while FL enables privacy-preserving training by aggregating model updates without transferring sensitive video data to centralized servers. The effectiveness of the proposed system is validated using the FaceForensics++ dataset under resource-constrained conditions. Experimental results demonstrate that FL-TENB4 achieves high detection accuracy, reduced model size, and low inference latency, making it highly suitable for real-world CCTV environments.

Overview

CoChat Summary

## TL;DR

FL-TENB4 is a federated-learning-enhanced TinyML deepfake detector built on EfficientNetB4-Lite for resource-constrained CCTV environments. On FaceForensics++ it reports 94.2% accuracy, 93.5% F1-score, 0.96 ROC-AUC, 12 ms per frame latency, 4.2 MB model size, and 150 mW inference energy.

## Key findings

  • On FaceForensics++ (reported as C0, C23, and C40), FL-TENB4 achieved 96.5% accuracy on C0, 94.2% on C23, and 89.8% on C40; the paper says C23 is most representative of real CCTV conditions.
  • The model reported 94.2% detection accuracy, 93.5% F1-score, 0.96 ROC-AUC, 12 ms per frame inference latency, 4.2 MB model size, and 150 mW energy consumption during inference.
  • Compared with listed baselines, FL-TENB4 outperformed Multi-View Inconsistency (91.8%) and CNN-SVM (90.2%), and was slightly above or comparable to XceptionNet (93.5%) in accuracy.
  • The paper states that FL-TENB4 reduces bandwidth requirements through quantization/sparsification/selective aggregation and supports scalable collaborative training across edge devices.

## Methods

  • The paper proposes a three-layer architecture: device layer, edge layer, and cloud layer, combining TinyML with Federated Learning (FedAvg) for privacy-preserving deepfake detection in CCTV video streams.
  • The local detector is EfficientNetB4-Lite optimized with quantization and pruning; frames are resized to 260 × 260, normalized, converted to tensors, and classified as real/fake in real time.
  • Experiments/simulations used FaceForensics++ and in-house CCTV data; the paper reports a 70:15:15 train/validation/test split and evaluation across C0, C23, and C40 compression levels, with comparisons to several existing methods.

## Limitations

  • The paper states that reliance on the FaceForensics++ dataset and a custom CCTV dataset constrains generalizability to real-world scenarios and a broader range of video qualities, compression levels, and manipulation types.
  • The paper notes that real-world CCTV footage often contains frames without human faces, while the model focuses on facial manipulations; it proposes a face-detection module as a remedy.
  • The text also makes evident that the FL framework does not explicitly address adversarial attacks such as model poisoning during aggregation; this is stated in the discussion, not as a formal limitation section.

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

Integrity signals

Registry

Crossref · 10.3390/s25030788

Cites

29 of 29 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Architecture overview.
Figure 1 Architecture overview.CC BY · PubMed Central
Figure 2 Simulation results of FL-TENB4 model.
Figure 2 Simulation results of FL-TENB4 model.CC BY · PubMed Central
Figure 3 Performance comparison.
Figure 3 Performance comparison.CC BY · PubMed Central
Figure 4 Comparison of efficiency metrics.
Figure 4 Comparison of efficiency metrics.CC BY · PubMed Central
Figure 5 Optimized bandwidth utilization and latency variability in Federated Learning with 100 edge devices.
Figure 5 Optimized bandwidth utilization and latency variability in Federated Learning with 100 edge devices.CC BY · PubMed Central

Verification

Claims the paper makes

4 Valid · 6 Not valid · 8 Not checked · 86 Own results and interpretations

34 more claims

Create a free account to see all 104 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 44
  • Probably not a substantial claim. 42

Create a free account to see these claims.

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.

1 of 1

Citations

Cites

Show all 29 references

Research works printed but not linked: 3

  1. Jeremiah, 2024, PrivacyGuard: Collaborative Edge-Cloud Computing Architecture for Attribute-Preserving Face Anonymization in CCTV Networks, Hum. Centric Comput. Inf. Sci., 14, 1
  2. Castro, 2023, Comprehensive Survey on AI-Based Technologies for Enhancing IoT Privacy and Security: Trends, Challenges, and Solutions, Hum. Centric Comput. Inf. Sci., 13, 39
  3. Kim, 2023, MBTI Personality Type Prediction Model Using WZT Analysis Based on the CNN Ensemble and GAN, Hum. Centric Comput. Inf. Sci., 13, 14

Cited by

See all 13

Metadata

Authors
Ha, J., El Azzaoui, A., & Park, J. H.
Affiliations
  • Seoul National University of Science and Technology (KR)
  • Department of Computer Science and Engineering, Seoul National University of Science and Technology, Seoul 01811, Republic of Korea
Corresponding author
Park, Jong Hyuk
Venue
Sensors
Publisher
MDPI AG
Published
2025-01-28
Volume
25
Issue
3
Pages
788
ISSN
1424-8220 (ISSN-L)
OpenAlex
W4406908477
Cited by
13
References
27
Citations by year
2026: 9 · 2025: 4
FWCI
4.52
Type
journal article
Language
en
Open access
Locations
  • Sensors (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
Korean government (2780000007)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
  • Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Face recognition and analysis
Keywords
deepfake detection, federated learning, TinyML, CCTV surveillance, edge computing, lightweight neural networks, privacy-preserving training, real-time video analysis, FaceForensics++

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 2026-09-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-09-29 (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

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: FL-TENB4: A Federated-Learning-Enhanced Tiny EfficientNetB4-Lite Approach for Deepfake Detection in CCTV Environments. https://app.cochat.ai/papers/10.3390/s25030788
The paper
Ha, J., El Azzaoui, A., & Park, J. H. (2025). FL-TENB4: A Federated-Learning-Enhanced Tiny EfficientNetB4-Lite Approach for Deepfake Detection in CCTV Environments. Sensors. https://doi.org/10.3390/s25030788

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.