Parameter-efficient fine-tuning (PEFT) is a machine-learning method for adapting a pre-trained model to a new task while training only a small set of parameters. Instead of updating the whole model, defenders add lightweight components or tune a limited subset of weights. This cuts compute, memory use, and storage needs, which makes adaptation practical for smaller teams and edge or offline deployments.
In cyber security, PEFT matters because it helps organizations customize AI systems for local languages, policies, and threat data without rebuilding a model from scratch. Security teams may use it to adapt detectors for phishing, fraud, malware analysis, or incident triage. The same technique can also help attackers cheaply modify open models to generate more convincing lures, automate translation, or evade simple content filters. For defense, the key controls are provenance, access control for training data, and testing for data leakage, bias, and prompt or model abuse.