PrivCore: Multiplication-activation co-reduction for efficient private inference.

Area of Science:

• Cryptography and Machine Learning
• Secure Computation
• Deep Learning Architectures

Background:

• Deep neural networks (DNNs) combined with secure 2-party computation (2PC) enable private inference (PI) but incur significant communication and latency costs.
• Previous research focused on optimizing non-linear operations, overlooking the substantial communication overhead from linear convolutions in PI protocols.

Purpose of the Study:

• To develop a framework, PrivCore, that jointly optimizes linear and non-linear DNN operators for efficient private inference.
• To significantly reduce communication and latency penalties in PI while maintaining inference accuracy.

Main Methods:

• PrivCore employs a co-design of sparse Winograd convolution and fine-grained activation reduction for optimized ciphertext computation.
• A two-tiered Winograd-aware structured pruning method is introduced to reduce multiplications by removing spatial filters and Winograd vectors.
• A sensitivity-based differentiable activation approximation and coefficient-adaptive polynomial replacement are used to automate ReLU selection and mitigate accuracy loss.

Main Results:

• PrivCore achieved a 2.2x communication reduction with 1.8% higher accuracy than SENet on CIFAR-100.
• On ImageNet, PrivCore demonstrated a 2.0x total communication reduction with equivalent accuracy compared to CoPriv.
• Experiments across various models and datasets consistently validate PrivCore's effectiveness in improving PI efficiency.

Conclusions:

• PrivCore offers a novel approach to significantly enhance the efficiency of private inference by optimizing both linear and non-linear operations.
• The framework successfully addresses the communication bottlenecks in PI protocols without compromising model accuracy.
• PrivCore represents a significant advancement in enabling practical and efficient privacy-preserving machine learning inference.