MetaKernel: Enabling Efficient Encrypted Neural Network Inference Through Unified MVM(…)
summary
Presents MKR, a compiler technique for CKKS fully homomorphic encryption that unifies and optimizes Matrix-Vector Multiplication and Convolution kernels for neural network inference. MKR decomposes kernels into composable MetaKernels and uses horizontal and vertical batching plus a rotation-aware cost model for data packing to cut expensive ciphertext rotations, improve slot utilization, support arbitrary input sizes and preserve output layout, and reports large kernel and end-to-end speedups over a state-of-the-art FHE compiler.
AI-written summary · about
related talks
44mIntel Math Kernel Library Perspectives & Latest Advances
37mOpen-source Machine Learning on Encrypted Data
38mHardware-based KMS Plug-in to Protect Secrets in Kubernetes
52mMachine Learning over Encrypted Data with Fully Homomorphic Encryption
21mKTransformers: Unleashing the Full Potential of CPU/GPU Hybrid Inference for M...
16mRegister Tiling for Unstructured Sparsity in Neural Network Inference
01h 06mObfuscation: Past, Present, and Possible Futures
23mKubernetes + Encrypted Memory = Security * Privacy