GRM-20242 CIPHER: Covert Influence Passed via Hidden Encoding in Representations Evaluating Subliminal Bias Transfer During Knowledge Distillation

GRM-20242 CIPHER: Covert Influence Passed via Hidden Encoding in Representations Evaluating Subliminal Bias Transfer During Knowledge Distillation

https://www.kennesaw.edu/ccse/events/computing-showcase/fa25-cday-program.php

AI models can inherit hidden behavioral biases when student models learn from teacher outputs during knowledge distillation. Project CIPHER investigates whether covert signals, such as zero-width Unicode characters or column order shifts, can transmit bias from a teacher model to a student model even when the student never receives group labels. Using an experimental pipeline with controlled subliminal cues and dual distillation, the project aims to reproduce and measure subtle bias transfer. Preliminary results showed that weak signals produce no measurable bias, while the redesigned high-frequency signal and MLP student architecture reveal quantifiable disparity.