Totem - Runtime behavioral integrity verification for LLMs in decision support systems via salted probing and cryptographic attestation

Read full text Summarise with VIRAT
Authors: Andrea Nucciarelli
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

Totem is a backend-agnostic proxy for runtime behavioral integrity verification of Large Language Models deployed as decision support systems. While existing LLM security tools protect against malicious users, none verify whether the model itself has been compromised. Totem fills this gap through three mechanisms: behavioral profiling via refusal classification, salted probes with steganographic triggers to prevent selective evasion, and cryptographically signed behavioral baselines (Model Manifest) authenticated via Ed25519 digital signatures. Evaluated across three model families and two attack vectors, Totem achieves 70% must-refuse detection rate on uncensored model swaps with 0% false positives, while naive baselines detect 0% of the same swaps. Released as open-source software under the Apache 2.0 license.

Fields of Study & Subjects

Adversarial Robustness in Machine Learning Security and Verification in Computing Physical Unclonable Functions (PUFs) and Hardware Security

Publication details

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