SocialGuard: An Explainable Multi-Signal Machine Learning Framework for Social Media Fake Account Detection

Abstract

Social media platforms are increasingly exposed to fake, spam, bot-operated and otherwise deceptive accounts that rule-based filters struggle to identify. This paper presents SocialGuard, an explainable multi-signal machine learning framework for fake-account detection. Instead of judging an account from a few profile fields, SocialGuard fuses four independent families of evidence: profile attributes, behavioral patterns, content characteristics and social-network structure. A classification layer that compares Logistic Regression, Random Forest, gradient-boosted trees and a neural network produces a fake-probability and a 0–100 risk score. An explainability layer built on Shapley-value attributions turns each score into evidence-based, feature-level reasoning, and a graph-analysis layer exposes suspicious clusters of accounts with unusually dense interaction. We formulate a research question and a four-step feature-ablation protocol that tests whether combining signal families improves detection over profile-only models. Because the evaluation on real benchmark data is still in progress, we validate the end-to-end pipeline on a controlled synthetic dataset of 3,000 accounts; those numbers demonstrate the methodology and must not be read as real-world performance. The framework is delivered as a web application with a REST API and an investigation dashboard that ends in a final investigation report.

Introduction

Social media platforms face a growing problem of fake, duplicate, spam, and bot-operated accounts. These accounts are used for scams, spam campaigns, artificial engagement, and spreading misleading information. Traditional manual reviews and rule-based systems are insufficient because of the large number of accounts and the constantly changing techniques used by attackers. Machine learning (ML) provides an automated way to detect suspicious accounts by identifying patterns in account data.

SocialGuard is a proposed AI-based framework designed to detect fake social media accounts more accurately and transparently. Unlike traditional systems that rely on a single classifier and provide only a fake-or-real result, SocialGuard combines four types of signals: profile, behavioral, content, and network features. It compares four machine learning models: Logistic Regression, Random Forest, gradient-boosted trees (XGBoost or LightGBM), and a feed-forward neural network.

The system addresses the limitations of existing research by providing three major capabilities:

• Multi-signal detection: Combines account information, user activity, posted content, and social-network relationships to identify suspicious patterns.

• Explainable AI: Generates a risk score from 0 to 100 and explains the prediction using supporting and contradicting evidence, with feature contributions calculated using SHAP.

• Suspicious cluster detection: Uses graph analysis, community detection, and network-density measurements to identify groups of accounts that may be coordinating their activities.

SocialGuard uses public datasets, including Kaggle datasets and the TwiBot-20 benchmark, where access and licensing permit. The data is processed into a unified format, and the models are evaluated using precision, recall, F1-score, ROC-AUC, PR-AUC, and confusion matrices. These metrics help assess detection quality, particularly when fake accounts are much less common than genuine ones.

The proposed system also addresses challenges such as evolving attacker behavior, incomplete and imbalanced datasets, data drift, scalability, privacy, and the need for understandable predictions. Its implementation uses Python-based machine learning and data-processing libraries, SHAP for explainability, NetworkX for graph analysis, FastAPI for backend services, PostgreSQL for data storage, and React for the investigation dashboard.

The research reviews existing studies from 2024–2025 and identifies gaps in combining all four signal types, explaining individual predictions, and detecting coordinated account groups. SocialGuard aims to address these gaps through a unified, explainable, and reproducible detection pipeline. Its evaluation protocol includes four feature-ablation experiments to measure the contribution of each signal family.

Conclusion

This paper presented SocialGuard, an explainable multi-signal framework that detects fake social media accounts by fusing profile, behavioral, content and network evidence, explains each prediction through a risk score and feature-level evidence, and reveals coordinated groups through cluster detection. A four-step ablation protocol was defined to test whether multi-signal fusion improves on profile-only models, and the pipeline was validated end to end on synthetic data. Future work will (i) build the unified dataset from Kaggle, TwiBot-20 and other public sources and run the experiments on real data with cross-validation and external testing, (ii) replace the stand-in components with XGBoost and SHAP as planned, (iii) add drift monitoring and periodic retraining, and (iv) implement the FastAPI backend, PostgreSQL storage and React investigation dashboard, with optional data adapters for legitimate platform APIs.

References

[1] J. S. Jothi, A. K. Akash, R. Akash, and M. Deepak, “Detection of fake profiles on social networking,” 2024. [2] R. Thangamani, M. Vimaladevi, J. Mohammed Rifath, G. Harini, and S. Hari, “Fake social media detection using machine learning,” 2024. [3] A. Padmavathi and K. B. Vaisshnavi, “Comparative analysis of fake account detection using machine learning algorithms,” 2024. [4] K. Bhargavi, P. M. D. Ali Khan, P. Imran Khan, and T. Sivakrishna, “Fake social media accounts detection using machine learning,” 2025. [5] S. Feng, H. Wan, N. Wang, J. Li, and M. Luo, “TwiBot-20: A comprehensive Twitter bot detection benchmark,” in Proc. 30th ACM Int. Conf. Information and Knowledge Management (CIKM), 2021. [6] L. Breiman, “Random forests,” Machine Learning, vol. 45, no. 1, pp. 5–32, 2001. [7] T. Chen and C. Guestrin, “XGBoost: A scalable tree boosting system,” in Proc. 22nd ACM SIGKDD Int. Conf. Knowledge Discovery and Data Mining, 2016, pp. 785–794. [8] S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” in Advances in Neural Information Processing Systems 30 (NeurIPS), 2017. [9] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre, “Fast unfolding of communities in large networks,” Journal of Statistical Mechanics: Theory and Experiment, vol. 2008, no. 10, P10008, 2008. [10] A. A. Hagberg, D. A. Schult, and P. J. Swart, “Exploring network structure, dynamics, and function using NetworkX,” in Proc. 7th Python in Science Conference (SciPy), 2008, pp. 11–15. [11] F. Pedregosa et al., “Scikit-learn: Machine learning in Python,” Journal of Machine Learning Research, vol. 12, pp. 2825–2830, 2011.

Copyright

Copyright © 2026 Prof. Pallavi A. Pathare, Krushna Jagtap, Aaditya Kulkarni, Rohit Gosavi, Harshal Gosavi, Prof. Suvarna S. Wakchaure, Dr. Rahul M. Dhokane. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.