Multi-Layer Honeypot Framework for Defending Pre-Trained Language Models Against Backdoor Attacks in Classification Tasks

Summary

The Multi-Layer Honeypot (MLH) framework is a defense for pre-trained language models against backdoor attacks in text classification. It deploys honeypot modules across the lower, middle and higher layers during fine-tuning, aggregates layer-wise anomaly scores, and down-weights highly anomalous samples in the task loss. Experiments on SST-2, SMS, AGNews and Amazon Reviews with BERT, RoBERTa, XLNet, OPT-1.3B and LLaMA-7B report reduced attack success rates while keeping classification performance.

Topics

• Adversarial machine learning

Related items

• MediumTrained to betray: How post-training supply chain attacks hijack coding agentsSimilar attack · Microsoft Security Blog
• MediumRewriting less than 10 neurons in an LLM can turn it into a sleeper backdoorSimilar attack · CSO Online
• InfoLazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI AgentsSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
• InfoHijacking a Trusted LLM in Real-World Quantized Deployments Only Requires a Disguised AdapterSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
• InfoPoison-then-Hide: Finetuning-Activated Backdoor Attack on Pretrained Vision EncodersSimilar attack · Arxiv (cs.RO + cs.CV security)