PAPER: Privacy-Preserving Convolutional Neural Networks using Low-Degree Polynomial Approximations and Structural Optimizations on Leveled FHE

The project addresses a critical bottleneck in Fully Homomorphic Encryption (FHE): the trade-off between bootstrapping frequency (which is slow) and multiplicative depth (which limits network complexity). By focusing on low-degree polynomial approximations for activations, it targets the 'Leveled' FHE niche. Quantitatively, the project has zero stars despite being 200+ days old, indicating it is currently a dormant or obscure academic reference implementation rather than a community-driven tool. While the mathematical approach is a 'novel combination' of structural optimization and approximation theory, its defensibility is low because the code lacks adoption, and the technique can be replicated by specialized FHE firms like Zama, Duality, or Microsoft Research once published. Frontier labs (OpenAI/Anthropic) have shown little interest in FHE, preferring Trusted Execution Environments (TEEs) or differential privacy, making the frontier risk low. However, the risk of platform domination by FHE-specific infrastructure providers (who could absorb these techniques into their compilers) is medium-to-high.