Скачать или смотреть Words Under Attack: Exposing Vulnerabilities in Text-driven Semantic Communication

Abstract: The advent of next-generation wireless networks, including 5G and beyond, coupled with the rapid proliferation of the Internet of Things (IoT), has intensified the demand for bandwidth by enabling a vast number of devices to transmit data concurrently. Traditional communication systems emphasize the reliable end-to-end delivery of symbols or bits without explicitly optimizing for the receiver-end application, which can result in inefficient bandwidth utilization. This limitation has motivated the development of semantic communication, an emerging paradigm that leverages large language models (LLMs) to transmit only task-relevant information rather than the entire message. By focusing on semantic content aligned with the receiver’s task, semantic communication substantially improves spectral efficiency and mitigates spectrum scarcity challenges in future networks. However, despite its potential, semantic communication introduces new security concerns. In particular, the reliance on deep learning models makes semantic systems highly susceptible to adversarial perturbations. While existing research has predominantly focused on the robustness of image-based semantic communication, the vulnerability of text-driven semantic communication remains underexplored. To address this gap, we present SemPerGe (Semantic Perturbation Generator), the first framework designed to generate targeted adversarial perturbations in text-based semantic communication. SemPerGe operates in a model-agnostic setting, requiring no prior knowledge of the model architecture, parameters, or outputs. It consists of two key components: (i) a Significant Token Marker to identify semantically critical tokens, and (ii) a Perturbation Generator that modifies these tokens to subtly alter semantics while preserving fluency and coherence. By exposing how easily transmitted meaning can be manipulated, SemPerGe reveals a critical and previously overlooked security vulnerability in semantic communication. The talk will demonstrate why securing semantic channels is essential for the reliability of next-generation networks and will outline future directions for building more resilient communication systems.

Speaker's Bio: Afia Anjum is a Ph.D. Candidate in the Computer Science department of the University of Texas at Arlington. In parallel with her doctoral studies, she works as a Graduate Researcher at Los Alamos National Laboratory (LANL) on a project supported by the Office of Cybersecurity, Energy Security, and Emergency Response under the U.S. Department of Energy. Her research focuses on designing mechanisms that enhance efficiency, resilience, and trust in wirelessly connected systems, with emphasis on data-centric architectures such as Named Data Networking, advanced wireless networks including 5G and future generations, and intent-centric paradigms like semantic communication.


About the Monthly Rising Star Symposium Series: The IEEE TCCN Special Interest Group for AI and Machine Learning in Security conducts a rising star symposium series where emerging scholars (e.g., senior PhD candidates and postdocs) present their research to a broader audience with the intention of fostering more mentorship, collaboration, and employment opportunities between the speakers and audience.