量化触发后门：跨量化器迁移性与验证-部署差距
原标题：Quantization-Triggered Backdoors in Language Models: Cross-Quantizer Transferability and the Validation--Deployment Gap
AI 摘要
该论文提出量化触发后门攻击，利用后训练量化作为触发器，使模型在源精度验证时表现良性，但在INT8或4位压缩后激活恶意行为。研究将攻击扩展到多语言编码器-解码器seq2seq模型，并发现攻击持久性取决于量化器几何和模型架构而非标称位宽。结果表明源精度审计不足以认证部署行为，最终部署配置必须纳入行为认证。
正文节选
Quantization-Triggered Backdoors in Language Models: Cross-Quantizer Transferability and the Validation–Deployment Gap Abstract Post-training quantization is often treated as a semantically neutral optimization for edge deployment of Large Language Models. When a full-precision source checkpoint is evaluated and quantization is applied downstream without equivalent re-evaluation, this workflow creates a structural validation–deployment gap: because quantization is a many-to-one mapping over para