Breaking the Layer Barrier - Remodeling Private Transformer Inference with Hybrid CKKS and MPC

Learn about an innovative framework for private Transformer inference that combines Homomorphic Encryption and Secure Multi-party Computation to protect data privacy in this 15-minute conference presentation. Discover how the proposed BLB (Breaking the Layer Barrier) framework addresses the significant communication costs introduced by conversions between HE and MPC in existing methods by breaking down layers into fine-grained operators and fusing adjacent linear operators. Explore the first secure conversion protocol between CKKS and MPC that enables CKKS-based computation of fused operators while managing increased ciphertext bit width. Examine the efficient matrix multiplication protocol designed specifically for fused computation in Transformers and review extensive evaluation results on BERT-base, BERT-large, and GPT2-base models that demonstrate BLB's superior performance with 21x reduction in communication overhead compared to BOLT and 2x reduction compared to Bumblebee, along with substantial latency improvements of 13x and 1.8x respectively when utilizing GPU acceleration.