Seed-Induced Uniqueness in Transformer Models: Subspace Alignment Governs Subliminal Transfer

TL;DR

This work questions the assumption that high global representational similarity drives subliminal transfer between Transformer models. By comparing same-base and independently initialized (different-base) students, the authors show that leakage is governed by alignment within a trait-discriminative subspace rather than overall CNN-like similarity, with same-base seeds producing substantially more leakage ($\tau \approx 0.24$) than different-base seeds ($\tau \approx 0.12$–$0.13$) even when $CKA>0.9$. They introduce a subspace-level CKA diagnostic and residualized probes to detect leakage more reliably, and demonstrate mitigation via a projection penalty, adversarial reversal, and right-for-the-wrong-reasons regularization that suppress trait-subspace alignment without harming public-task fidelity. The results argue for seed-aware deployment strategies and subspace diagnostics to strengthen secure multi-model systems, particularly in federated or adversarial contexts.

Abstract

We analyze subliminal transfer in Transformer models, where a teacher embeds hidden traits that can be linearly decoded by a student without degrading main-task performance. Prior work often attributes transferability to global representational similarity, typically quantified with Centered Kernel Alignment (CKA). Using synthetic corpora with disentangled public and private labels, we distill students under matched and independent random initializations. We find that transfer strength hinges on alignment within a trait-discriminative subspace: same-seed students inherit this alignment and show higher leakage {τ\approx} 0.24, whereas different-seed students -- despite global CKA > 0.9 -- exhibit substantially reduced excess accuracy {τ\approx} 0.12 - 0.13. We formalize this with subspace-level CKA diagnostic and residualized probes, showing that leakage tracks alignment within the trait-discriminative subspace rather than global representational similarity. Security controls (projection penalty, adversarial reversal, right-for-the-wrong-reasons regularization) reduce leakage in same-base models without impairing public-task fidelity. These results establish seed-induced uniqueness as a resilience property and argue for subspace-aware diagnostics for secure multi-model deployments.

Figures (1)

• Figure 1: $\tau_{\text{resid}}$ vs. subspace CKA across ablation conditions.