Defending Against Text-Based Social Engineering Attacks Using Federated Adversarial Learning

Emdadul Haque Iram, Navid Nahiyan, Musfika Jahan, Md. Nazmul Hasan
Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments, PMLR 319:26-36, 2026.

Abstract

Text-based social engineering attacks such as phishing emails and scam messages have remained quite successful because language and obfuscation patterns are constantly adapted by adversaries, and centralised detection approaches raise serious privacy concerns as well as insufficient real-world deployment. This paper proposes a privacy-preserving and robust detection system combining Joint Embedding Predictive Architecture (JEPA) representation learning with federated learning, enabling multiple clients to collaboratively train a global model without exchanging raw user data. To overcome modelling capacity challenges in heterogeneous (non-IID) client distributions, a Mixture-of-Experts (MoE) design and a Kolmogorov–Arnold Network (KAN)-based prediction head are adopted. The global model is trained via iterative local optimisation and server-side aggregation, with robustness induced by a federated adversarial learning stage. Experimental results demonstrate that the JEPA-Federated-MoE/KAN pipeline consistently exhibits excellent detection performance with privacy preservation and enhanced flexibility against adversarial changes.

Cite this Paper

BibTeX
@InProceedings{pmlr-v319-iram26a, title = {Defending Against Text-Based Social Engineering Attacks Using Federated Adversarial Learning}, author = {Iram, Emdadul Haque and Nahiyan, Navid and Jahan, Musfika and Hasan, Md. Nazmul}, booktitle = {Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments}, pages = {26--36}, year = {2026}, editor = {Folorunso, Sakinat and Ogundokun, Roseline and Oladipo, Francisca}, volume = {319}, series = {Proceedings of Machine Learning Research}, month = {11--14 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v319/main/assets/iram26a/iram26a.pdf}, url = {https://proceedings.mlr.press/v319/iram26a.html}, abstract = {Text-based social engineering attacks such as phishing emails and scam messages have remained quite successful because language and obfuscation patterns are constantly adapted by adversaries, and centralised detection approaches raise serious privacy concerns as well as insufficient real-world deployment. This paper proposes a privacy-preserving and robust detection system combining Joint Embedding Predictive Architecture (JEPA) representation learning with federated learning, enabling multiple clients to collaboratively train a global model without exchanging raw user data. To overcome modelling capacity challenges in heterogeneous (non-IID) client distributions, a Mixture-of-Experts (MoE) design and a Kolmogorov–Arnold Network (KAN)-based prediction head are adopted. The global model is trained via iterative local optimisation and server-side aggregation, with robustness induced by a federated adversarial learning stage. Experimental results demonstrate that the JEPA-Federated-MoE/KAN pipeline consistently exhibits excellent detection performance with privacy preservation and enhanced flexibility against adversarial changes.} }
Endnote
%0 Conference Paper %T Defending Against Text-Based Social Engineering Attacks Using Federated Adversarial Learning %A Emdadul Haque Iram %A Navid Nahiyan %A Musfika Jahan %A Md. Nazmul Hasan %B Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments %C Proceedings of Machine Learning Research %D 2026 %E Sakinat Folorunso %E Roseline Ogundokun %E Francisca Oladipo %F pmlr-v319-iram26a %I PMLR %P 26--36 %U https://proceedings.mlr.press/v319/iram26a.html %V 319 %X Text-based social engineering attacks such as phishing emails and scam messages have remained quite successful because language and obfuscation patterns are constantly adapted by adversaries, and centralised detection approaches raise serious privacy concerns as well as insufficient real-world deployment. This paper proposes a privacy-preserving and robust detection system combining Joint Embedding Predictive Architecture (JEPA) representation learning with federated learning, enabling multiple clients to collaboratively train a global model without exchanging raw user data. To overcome modelling capacity challenges in heterogeneous (non-IID) client distributions, a Mixture-of-Experts (MoE) design and a Kolmogorov–Arnold Network (KAN)-based prediction head are adopted. The global model is trained via iterative local optimisation and server-side aggregation, with robustness induced by a federated adversarial learning stage. Experimental results demonstrate that the JEPA-Federated-MoE/KAN pipeline consistently exhibits excellent detection performance with privacy preservation and enhanced flexibility against adversarial changes.
APA
Iram, E.H., Nahiyan, N., Jahan, M. & Hasan, M.N.. (2026). Defending Against Text-Based Social Engineering Attacks Using Federated Adversarial Learning. Proceedings of IndabaX Nigeria 2026: Building Scalable AI That Works: From Research to Deployment in Resource-Constrained Environments, in Proceedings of Machine Learning Research 319:26-36 Available from https://proceedings.mlr.press/v319/iram26a.html.

Related Material