lorenzorovida/LowMemoryFHEResNet20

LowMemoryFHEResNet20 is a specialized research artifact providing a concrete implementation of the techniques described in its namesake paper. With 77 stars and 25 forks, it has served as a valuable reference for the Privacy-Preserving Machine Learning (PPML) community, particularly those working with Microsoft SEAL. However, its defensibility is limited: it is a static codebase (0.0 velocity) tied to a specific, aging model architecture (ResNet-20). The project's primary moat is the deep domain expertise required to manually optimize FHE parameters and ciphertext rotations to fit within specific memory constraints—a 'dark art' in cryptography. However, this moat is being rapidly eroded by the rise of FHE compilers like Zama's Concrete-ML and OpenMined's TenSEAL, which aim to automate these optimizations for any arbitrary model. Frontier labs are unlikely to compete here as FHE remains too computationally expensive for LLM-scale inference, and cloud providers (AWS/Azure) favor Trusted Execution Environments (TEEs) like Nitro or SGX for privacy-preserving compute due to their lower overhead. The highest risk is displacement by general-purpose FHE frameworks that will eventually render manual, model-specific optimizations like those in this repo obsolete within 1-2 years. It remains a solid reference for academic benchmarking but lacks the trajectory of a sustainable software project.