Mitigating the Impact of Adversarial Fine-tuning on Large Language Models through Perturbation Attenuation

Abstract The increasing complexity and scale of machine learning models have rendered them vulnerable to adversarial manipulation, particularly during fine-tuning processes where subtle perturbations can lead to significant degradation in model performance. Addressing this vulnerability, the proposed approach introduces a novel perturbation attenuation mechanism that integrates noise injection and gradient clipping during training, offering a substantial improvement in model resilience against adversarial attacks. Through a detailed evaluation of the Mistral Large model, the research demonstrates how the application of such attenuation techniques preserves model accuracy while mitigating the adverse effects of adversarial fine-tuning, significantly enhancing robustness without imposing excessive computational overhead. Experiments comparing the baseline model, adversarially fine-tuned model, and the model equipped with the proposed attenuation mechanism reveal that the latter successfully restores stability across various tasks, confirming the efficacy of the strategy in real-world settings. This work contributes to the growing need for robust defenses against adversarial inputs, illustrating how targeted interventions can maintain the integrity and reliability of machine learning systems even in adversarial environments. The findings offer valuable insights into how to counteract adversarial strategies in machine learning models while preserving their performance, providing a practical framework for improving model robustness in critical applications.