Method for enhancing the robustness of large language models against adversarial attacks

Abstract

The paper presents a mathematically grounded method for enhancing the robustness of large language models (LLMs) against input-level manipulations, systematizing adversarial attack vectors and mechanisms for their preventive neutralization. The introductory section outlines the fundamental architectural vulnerability of autoregressive transformers – the conflation of operational instructions and unverified data within a single memory space, which generates systemic compromise risks and forms the core of threats such as prompt injection. The main part of the study proposes a comprehensive apparatus that operates at the fundamental level of tensor mechanics. To prevent gradient merging, a geometric projection module is developed, which orthogonally separates the query and RAG–context embeddings. To eliminate the risk of systemic memory liquidity depletion, 4–bit asymmetric quantization of transformation matrices is applied. In parallel, a dynamic entropy adapter is introduced, which algorithmically adjusts the penalty coefficient variable at the pre–softmax generation stage, eliminating the risk of over–refusal of legitimate queries. Additionally, an extended comparative analysis is conducted, which summarizes the advantages of the proposed architecture over existing reactive methods and resource-intensive shadow models. It is mathematically proven that the proposed logit shift renders the optimization gradient degenerate, guaranteeing an exponential decay in the probability of a successful attack. The conclusions summarize the key results and outline promising directions for further research: development of formal security guarantees for multimodal environments, standardization of benchmarks, and evolution of comprehensive monitoring systems.