Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FakeZero: Real-Time, Privacy-Preserving Misinformation Detection for Facebook and X

Soufiane Essahli, Oussama Sarsar, Ahmed Bentajer, Anas Motii, Imane Fouad

TL;DR

FakeZero demonstrates that real-time misinformation detection can be executed entirely within a web browser for Facebook and X, achieving near-state-of-the-art accuracy while preserving user privacy through on-device inference. The approach combines compact encoder backbones (DistilBERT and TinyBERT), a three-stage curriculum, focal loss, and post-training quantization to deliver fast, memory-efficient, and privacy-preserving performance. It provides transparent runtime metrics (latency, memory, and model size) and a practical deployment blueprint for cross-platform, privacy-first misinformation tools. The work has potential impact for policymakers and researchers by enabling large-scale in-the-wild data collection with user consent, and sets a foundation for multilingual and multimodal future extensions in privacy-conscious settings.

Abstract

Social platforms distribute information at unprecedented speed, which in turn accelerates the spread of misinformation and threatens public discourse. We present FakeZero, a fully client-side, cross-platform browser extension that flags unreliable posts on Facebook and X (formerly Twitter) while the user scrolls. All computation, DOM scraping, tokenization, Transformer inference, and UI rendering run locally through the Chromium messaging API, so no personal data leaves the device. FakeZero employs a three-stage training curriculum: baseline fine-tuning and domain-adaptive training enhanced with focal loss, adversarial augmentation, and post-training quantization. Evaluated on a dataset of 239,000 posts, the DistilBERT-Quant model (67.6 MB) reaches 97.1% macro-F1, 97.4% accuracy, and an AUROC of 0.996, with a median latency of approximately 103 ms on a commodity laptop. A memory-efficient TinyBERT-Quant variant retains 95.7% macro-F1 and 96.1% accuracy while shrinking the model to 14.7 MB and lowering latency to approximately 40 ms, showing that high-quality fake-news detection is feasible under tight resource budgets with only modest performance loss. By providing inline credibility cues, the extension can serve as a valuable tool for policymakers seeking to curb the spread of misinformation across social networks. With user consent, FakeZero also opens the door for researchers to collect large-scale datasets of fake news in the wild, enabling deeper analysis and the development of more robust detection techniques.

FakeZero: Real-Time, Privacy-Preserving Misinformation Detection for Facebook and X

TL;DR

FakeZero demonstrates that real-time misinformation detection can be executed entirely within a web browser for Facebook and X, achieving near-state-of-the-art accuracy while preserving user privacy through on-device inference. The approach combines compact encoder backbones (DistilBERT and TinyBERT), a three-stage curriculum, focal loss, and post-training quantization to deliver fast, memory-efficient, and privacy-preserving performance. It provides transparent runtime metrics (latency, memory, and model size) and a practical deployment blueprint for cross-platform, privacy-first misinformation tools. The work has potential impact for policymakers and researchers by enabling large-scale in-the-wild data collection with user consent, and sets a foundation for multilingual and multimodal future extensions in privacy-conscious settings.

Abstract

Social platforms distribute information at unprecedented speed, which in turn accelerates the spread of misinformation and threatens public discourse. We present FakeZero, a fully client-side, cross-platform browser extension that flags unreliable posts on Facebook and X (formerly Twitter) while the user scrolls. All computation, DOM scraping, tokenization, Transformer inference, and UI rendering run locally through the Chromium messaging API, so no personal data leaves the device. FakeZero employs a three-stage training curriculum: baseline fine-tuning and domain-adaptive training enhanced with focal loss, adversarial augmentation, and post-training quantization. Evaluated on a dataset of 239,000 posts, the DistilBERT-Quant model (67.6 MB) reaches 97.1% macro-F1, 97.4% accuracy, and an AUROC of 0.996, with a median latency of approximately 103 ms on a commodity laptop. A memory-efficient TinyBERT-Quant variant retains 95.7% macro-F1 and 96.1% accuracy while shrinking the model to 14.7 MB and lowering latency to approximately 40 ms, showing that high-quality fake-news detection is feasible under tight resource budgets with only modest performance loss. By providing inline credibility cues, the extension can serve as a valuable tool for policymakers seeking to curb the spread of misinformation across social networks. With user consent, FakeZero also opens the door for researchers to collect large-scale datasets of fake news in the wild, enabling deeper analysis and the development of more robust detection techniques.

Paper Structure

This paper contains 32 sections, 4 equations, 8 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Browser–integrated detectors
  4. Transformer-centric text models
  5. Graph-, multimodal- and retrieval-augmented methods
  6. Surveys, Benchmarks, and Data Resources
  7. Remaining gaps and our contribution
  8. System Overview
  9. Datasets & Pre-processing
  10. Model Architecture & Training
  11. Encoder Selection and Baselines
  12. Training Curriculum and Enhancements
  13. Curriculum.
  14. Focal loss.
  15. Optimisation.
  16. ...and 17 more sections

Figures (8)

  • Figure 1: Client-side FakeZero pipeline from data preprocessing to in-browser fake content flagging.
  • Figure 2: Three-stage curriculum used to train FakeZero. Stage 0 warms up on balanced long-form news (ISOT); Stage 1 adapts to short, highly imbalanced political claims (LIAR); Stage 2 fine-tunes on a harder mix of FakeNewsNet (50 %), TruthSeeker (30 %), and PHEME (20 %). Posts not chosen for Stage 2 are split 1 : 1 into Dev and Test, preserving the overall 64 % misinformation prevalence and eliminating topic leakage.
  • Figure 3: Latency–accuracy Pareto frontier for all quantised checkpoints. Bubble area encodes ONNX size.
  • Figure 4: Best hyperparameter configurations
  • Figure 5: Confusion matrix for the test split.
  • ...and 3 more figures