Research Papers 论文研究 5h ago Updated 52m ago 更新于 52分钟前 47

Quantization-Triggered Backdoors in Language Models: Cross-Quantizer Transferability and the Validation--Deployment Gap 量化触发的语言模型后门:跨量化器可迁移性与验证-部署差距

Post-training quantization is not semantically neutral; it creates a structural validation-deployment gap where full-precision certification does not guarantee behavioral equivalence after compression The authors formalize this gap through Quantization Behavioral Equivalence Classes (QBECs) and prove that QBEC membership does not imply behavioral equivalence, establishing a theoretical foundation for quantization-triggered backdoor attacks A three-stage adversarial fine-tuning framework embeds l 后训练量化并非语义中性操作,源精度认证无法保证部署配置的行为等价性,存在结构性"验证-部署差距" 提出量化行为等价类(QBECs)理论框架,证明QBEC成员资格不蕴含行为等价性,为量化触发的后门攻击提供理论基础 通过三阶段对抗性微调框架,将潜伏恶意载荷嵌入满足源精度检查的模型,在INT8或4位压缩后激活对抗性行为 实验显示后门翻译模型从FP16的零友敌腐败率飙升至量化后85.02%的反转,政治立场分类器检测到高达ΔBias=0.33的意识形态偏移 跨量化器可转移性分析表明攻击持久性因量化方案和模型架构而异,而非仅由名义位宽决定

62
Hot 热度
72
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • Post-training quantization is not semantically neutral; it creates a structural validation-deployment gap where full-precision certification does not guarantee behavioral equivalence after compression
  • The authors formalize this gap through Quantization Behavioral Equivalence Classes (QBECs) and prove that QBEC membership does not imply behavioral equivalence, establishing a theoretical foundation for quantization-triggered backdoor attacks
  • A three-stage adversarial fine-tuning framework embeds latent malicious payloads that pass source-precision checks but activate targeted adversarial behavior upon INT8 or 4-bit compression
  • In tactical machine translation, backdoored models achieve up to 85.02% friend-foe inversion after quantization, and in political content analysis, an ideological shift of up to ΔBias=0.33 is measured
  • Cross-quantizer transferability analysis reveals attack persistence varies across quantization schemes and model architectures, not merely by nominal bit-width

Why It Matters

This research exposes a critical security vulnerability in the common ML deployment pipeline where models are validated at full precision and then quantized for edge deployment without re-evaluation. For AI practitioners and security researchers, it demonstrates that source-precision auditing alone is insufficient to guarantee trustworthy behavior in deployed models, fundamentally challenging assumptions about quantization as a benign optimization step.

Technical Details

  • The paper introduces Quantization Behavioral Equivalence Classes (QBECs) as a formal framework, proving that models within the same QBEC (i.e., mapping to the same quantized parameters) do not necessarily exhibit behavioral equivalence, thereby establishing the theoretical basis for quantization-triggered backdoors
  • A three-stage adversarial fine-tuning framework is developed to embed latent malicious payloads into models that satisfy full-precision evaluation checks while remaining dormant until INT8 or 4-bit compression activates the adversarial behavior
  • The threat is evaluated across two operationally motivated scenarios: tactical machine translation (friend-foe classification corruption) and political content analysis (ideological stance manipulation), extending prior backdoor research from decoder-only causal LMs to multilingual encoder-decoder sequence-to-sequence architectures
  • Quantification of impact shows backdoored translation models moving from 0% friend-foe corruption at repaired FP16 to up to 85.02% inversion post-quantization, with a paired stance classifier measuring ideological shifts up to ΔBias=0.33 after compression
  • Cross-quantizer transferability analysis demonstrates that attack persistence is not uniformly determined by nominal bit-width but varies significantly across different quantization schemes and model architectures

Industry Insight

  • Organizations deploying quantized models for edge inference must incorporate post-quantization behavioral certification into their security pipelines, as full-precision audits alone cannot detect quantization-triggered backdoors
  • The finding that attack transferability varies by quantization scheme and architecture—not just bit-width—suggests defense strategies should be quantizer-aware rather than relying on bit-width thresholds as a security proxy
  • The validation-deployment gap identified here should prompt a reevaluation of MLOps workflows, particularly in high-stakes domains like defense-related machine translation and political content moderation, where quantization-induced behavior shifts could have severe operational consequences

TL;DR

  • 后训练量化并非语义中性操作,源精度认证无法保证部署配置的行为等价性,存在结构性"验证-部署差距"
  • 提出量化行为等价类(QBECs)理论框架,证明QBEC成员资格不蕴含行为等价性,为量化触发的后门攻击提供理论基础
  • 通过三阶段对抗性微调框架,将潜伏恶意载荷嵌入满足源精度检查的模型,在INT8或4位压缩后激活对抗性行为
  • 实验显示后门翻译模型从FP16的零友敌腐败率飙升至量化后85.02%的反转,政治立场分类器检测到高达ΔBias=0.33的意识形态偏移
  • 跨量化器可转移性分析表明攻击持久性因量化方案和模型架构而异,而非仅由名义位宽决定

为什么值得看

本研究揭示了边缘AI部署中一个被忽视的安全漏洞:量化过程可能触发潜伏的后门攻击,这对AI系统的可信部署具有重大安全意义。研究将后门攻击从纯解码器因果LM扩展到多语言编码器-解码器序列到序列模型,拓展了攻击场景的适用范围。

技术解析

  • 理论框架:提出量化行为等价类(QBECs)概念,形式化验证-部署差距,证明量化作为参数空间的多对一映射,源精度认证无法保证部署配置的行为等价性。
  • 攻击方法:采用三阶段对抗性微调框架,在模型满足源精度检查的同时嵌入潜伏恶意载荷,使其在INT8或4位压缩后激活目标对抗性行为。
  • 实验场景:在战术机器翻译和政治内容分析两个操作驱动场景中评估威胁,扩展了先前仅在解码器-only因果LM上的工作。
  • 量化方案:测试了INT8和4-bit压缩方案,跨量化器可转移性分析显示攻击持久性因量化方案和模型架构而异。
  • 评估指标:友敌腐败率(inversion rate)和配对立场分类器的意识形态偏移(ΔBias)。

行业启示

  • 边缘AI部署的安全审计必须将最终部署配置纳入行为认证范围,仅依赖源精度审计无法排除量化触发的行为异常。
  • AI供应链安全需要建立端到端的验证流程,量化作为部署环节不应被视为"透明"操作,应纳入安全评估体系。
  • 模型压缩与部署流程中应引入量化感知的对抗性鲁棒性测试,特别是在军事、政治等高风险应用场景中。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Security 安全 Quantization 量化 LLM 大模型 Deployment 部署 Research 科学研究