Sparse Logit-Lens KL Divergence for LLM Backdoor Detection

Read full text Summarise with VIRAT
Authors: Ying Cong Seah
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Weight editing enables efficient modification of large language model behavior, but also provides a vector for implanting backdoors that are difficult to detect using standard techniques. We develop a general-purpose methodology for analyzing weight-edited backdoors based on sparse logit-lens KL divergence---comparing intermediate-layer predictions between modified and reference models using sparse top-k probability distributions and raw (non-chat-template) tokenization. We validate this methodology on four backdoored models spanning two architectures and two scales: a Qwen 2.5 7B model and three DeepSeek V3 671B Mixture-of-Experts models from the Jane Street "Dormant Model" puzzle. Our analysis reveals structural properties that may characterize a broader class of weight-edited backdoors: (1) rank-1 dominance---removing each layer's top singular value from the weight delta eliminates 95% of trigger-induced divergence; (2) progressive detection---trigger recognition builds cumulatively across layers proportional to model depth; and (3) scale invariance---the mechanism is structurally identical between 7B and 671B. We further show that eight commonly attempted detection approaches fail on this class of backdoor. Our methodology detected backdoor signatures at the 671B scale for a total compute cost of $66.

Fields of Study & Subjects

Adversarial Robustness in Machine Learning Digital Media Forensic Detection Physical Unclonable Functions (PUFs) and Hardware Security

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
30 March 2026
Language
EN
Document type
Preprint