Reconstructing Trust Embeddings from Siamese Trust Scores

This paper dives into reversing the process of turning basic one-dimensional trust scores from security frameworks back into detailed high-dimensional embeddings that represent device trustworthiness, proposing a straightforward method that stitches together paired scores with statistical moments and proves it converges to a unique solution using fixed-point math. Through simulations with noisy data on 20 devices over 10 time steps, it shows the reconstructions keep the original geometric relationships intact, backed by error guarantees tied to data length, but warns that sharing these scores poses a real privacy threat by exposing underlying behaviors of devices and models. To counter this, it suggests fixes like rounding scores, injecting controlled noise, or scrambling embeddings, all while weighing the trade-offs between transparency and secrecy in connected AI setups.