Date: 12/06/2026
Hour: 14.00
Student Name & ID: Bora Akpınar & 245108106
Supervisor: Asst. Prof. Metin Öztürk
Topic: KB-Aware Adversarial Semantic Jamming and Knowledge-Grounded LLM Decoding for Military UAV Communications
Location: A-424
Abstract: Semantic communication has emerged as a promising paradigm for next-generation military communication systems, enabling efficient transmission of meaningful information rather than raw data. In mission-critical UAV-to-command scenarios, the integrity of semantically encoded messages is paramount, as adversarial manipulation at the meaning layer can lead to catastrophic decision errors without triggering classical bit-level error detection mechanisms. This study introduces a novel four-level adversarial semantic jamming model, including a KB-aware single-token attack (S3-single) and a coordinated multi-token attack (S3-multi), where the adversary exploits knowledge of the military knowledge base to craft KB-consistent yet maximally decision-disrupting substitutions. To counter these threats, a knowledge-grounded LLM decoder architecture is proposed, in which constrained generation prevents hallucination at a structural level. The system is evaluated through a four-way ablation framework comparing classical error correction, RAG-only, unconstrained LLM, and the proposed KB-LLM decoder. A novel Decision Accuracy metric is introduced to measure whether repaired messages yield correct military decisions, complementing standard semantic similarity scores. Additionally, a token vulnerability analysis investigates which triplet element — entity, threat level, distance, or location — is most critical to decision accuracy under adversarial conditions. Experimental results are expected to demonstrate that KB-grounded LLM decoding significantly outperforms baseline systems under adversarial semantic jamming, with measurable breaking-point thresholds.