Reproduce “POSEIDON: Privacy-Preserving Federated Neural Network Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sinem Sav, Apostolos Pyrgelis, Juan R. Troncoso-Pastoriza, David Froelicher, Jean-Philippe Bossuat, Joao Sa Sousa, Jean-Pierre Hubaux
POSEIDON is the first system to enable quantum-resistant, privacy-preserving federated neural network training and inference in an N-party setting, utilizing multiparty lattice-based homomorphic encryption. It facilitates collaborative AI without centralizing data, achieving model accuracy comparable to non-private methods while significantly speeding up training (e.g., MNIST in less than 1.4 hours for 10 parties) and providing robust N-1 passive adversary security.
Key insights
- Multiparty homomorphic encryption can be practically scaled for complex neural network training and inference, offering strong N-1 passive adversary security and post-quantum resistance.
- Optimized data packing strategies like Alternating Packing significantly reduce computational overhead by minimizing expensive homomorphic rotations, enabling logarithmic scaling with features and neurons.
- Embedding arbitrary linear transformations directly into the distributed bootstrapping process can effectively perform complex operations 'for free' during ciphertext refreshing, boosting efficiency for CNNs and other models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.