LLM Security And Guardrail Defense Techniques In Web Applications
DOI:
https://doi.org/10.56472/ICCSAIML25-127Keywords:
Adversarial Attacks, AI Security, Data Privacy, Guardrails, Large Language Models (LLMs), LLM Security, Model Integrity, Robustness in LLMs, Safety Constraints, Security Threats in LLMs, Training Data SecurityAbstract
Adversarial attacks pose an important threat to the security and trustworthiness of large language models (LLMs). These models are vulnerable to carefully engineered input designed to exploit their weaknesses, degrade system performance, and extract private information. Practical countermeasures need a strong offensive strategy that includes adversarial scenarios simulation to test how models hold up in various conditions. Adversarial data poisoning and evasion techniques are particularly useful as they can reveal strengths during training and inference processes, respectively. Through systematic identification and mitigation of every vulnerability, organizations can strengthen the model's robustness to real-world attacks. This technical framework also emphasizes the need to embed adversarial simulations into the security cycle of LLMs to prevent risks associated with malicious actors. Via iterative analysis, an organization can increase the robustness of a model and build a resilient infrastructure for secure implementation of LLMs in sensitive/high-level environments
Downloads
References
[1] E. Dorsey, “Antitrust in Retrograde: The Consumer Welfare Standard, Socio-Political Goals, and the Future of Enforcement,” SSRN Electronic Journal, 2020, doi: 10.2139/ssrn.3733666.
[2] S. Jing, Y. Liu, D. Liu, and J. Guo, “Research on a New Synthesis of LLM-105 Using N-Nitroso-bis(cyanomethyl)amine,” Central European Journal of Energetic Materials, vol. 13, no. 1, pp. 21–32, 2016, doi: 10.22211/cejem/64962.
[3] L. You, Y. Li, Y. Wang, J. Zhang, and Y. Yang, “A deep learning-based RNNs model for automatic security audit of short messages,” in 2016 16th International Symposium on Communications and Information Technologies (ISCIT), IEEE, Sep. 2016, pp. 225–229.
[4] X. Pan, M. Zhang, S. Ji, and M. Yang, “Privacy Risks of General-Purpose Language Models,” in 2020 IEEE Symposium on Security and Privacy (SP), IEEE, May 2020, pp. 1314–1331.: https://doi.org/10.1109/sp40000.2020.00095
[5] B. Settles, G. T. LaFlair, and M. Hagiwara, “Machine Learning–Driven Language Assessment,” Transactions of the Association for Computational Linguistics, vol. 8, pp. 247–263, Dec. 2020, doi: 10.1162/tacl_a_00310.
[6] H. Chen, B. D. Rouhani, C. Fu, J. Zhao, and F. Koushanfar, “Deepmarks: A secure fingerprinting framework for digital rights management of deep learning models,” in Proceedings of the 2019 on International Conference on Multimedia Retrieval, New York, NY, USA: ACM, Jun. 2019, pp. 105–113.
[7] A. van den Berghe, R. Scandariato, K. Yskout, and W. Joosen, “Design notations for secure software: a systematic literature review,” Software & Systems Modeling, vol. 16, no. 3, pp. 809–831, Aug. 2015, doi: 10.1007/s10270-015-0486-9.
[8] L. Song, R. Shokri, and P. Mittal, “Privacy Risks of Securing Machine Learning Models against Adversarial Examples,” in Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, New York, NY, USA: ACM, Nov. 2019, pp. 241–257.
