The Root Causes and Defense Methods of the Security Degradation of Large-Scale Models after Fine-Tuning

After undergoing safety alignment, large language models may still lose their safety safeguards during third-party fine-tuning, resulting in a degradation of safety. Previous studies have advanced rapidly on both the offensive and defensive sides. However, the defensive effectiveness is limited by training parameters and attack patterns, lacks a unified explanation, and is difficult to support pre-event risk control. Therefore, this paper traces the evolution from explicit harmful fine-tuning to benign data-induced degradation attacks, with the characterization of root causes of similarity as the main thread. The attacks are classified into an evolving process with decreasing thresholds, and the defenses are categorized according to the intervention positions as alignment reinforcement, fine-tuning constraints, and post-processing fixes, etc. The evidence and failure conditions of each route's representative work under the experimental settings are compared, and the relationship between the characterization of similarity and security degradation is further analyzed. The review analysis results indicate that the attack methods are evolving from explicit harmful data to covert and benign forms. The current defense effectiveness is still significantly influenced by the training configuration and the form of the attack data. The degree of similarity in representation between the upstream safety-aligned data and the downstream fine-tuning data is significantly correlated with the degree of safety degradation. Therefore, this indicator is expected to be used for risk assessment before fine-tuning and to provide quantifiable prior information for the subsequent design of defense mechanisms.