Work

Conditional Subliminal Learning: Gated, Cross-Domain, and Mechanistically Explained Transfer of a Hidden Encryption-Algorithm Preference

Year: 2026

Type: preprint

Abstract: This paper extends subliminal learning (Cloud et al., 2026) to the security domain. We demonstrate that a hidden preference for the DES encryption algorithm can be transferred from a biased teacher model to a student model through semantically unrelated fine-tuning data. We introduce three contributions: (1) a gated transfer mechanism activated only by a legal-register stylistic trigger, (2) cross-domain leakage into generated Python code, and (3) a mechanistic explanation via the conditional D_diff diagnostic. The work surfaces a novel supply-chain risk in LLM fine-tuning pipelines.

Source: Zenodo (CERN European Organization for Nuclear Research)

DOI: 10.5281/zenodo.23940082

Author: Rishabh Kumar

Institution: Rheinland-Pfälzische Technische Universität Kaiserslautern-Landau

Language: English

Topic: Adversarial Robustness in Machine Learning

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: green

OpenAlex ID: W7225240631

Cites:

Cited by:

Related to: