NetDP: In-Network Differential Privacy for Large-Scale Data Processing

doi:10.1109/TGCN.2024.3432781

NetDP: In-Network Differential Privacy for Large-Scale Data Processing

2024-09-01

0

PRE

AI

Muhammad Khurram Khan

DOI：10.1109/TGCN.2024.3432781

Original

Original request for help

Save

Abstract

En 中文

Radio access network (RAN) enables large-scale collection of sensitive data. Privacy-preserving techniques aim to learn knowledge from sensitive data to improve services without compromising privacy. However, as the data scale increases, enforcing privacy-preserving techniques on sensitive data may consume a considerable amount of system resources and impose performance penalties. To reduce system resource consumption, we present NetDP, an in-network architecture for privacy-preserving techniques by leveraging programmable switches to improve resource efficiency (i.e., CPU cycles, network bandwidth, and privacy budgets). The key idea of NetDP is to accommodate and exploit cryptographic operators to reduce resource consumption rather than repetitively and exhaustively suppressing the impact of these techniques. To the best of our knowledge, this is the first time that privacy-preserving techniques in a large-scale data processing system have been enforced on programmable switches. Our experiments based on Tofino switches indicate that NetDP significantly reduces computation latency (e.g., 40.2%-55.8% latency in computations) without impacting fidelity.

Keywords:

Privacy

Noise

Differential privacy

Data processing

Sensitivity

Computer architecture

Pipelines

In-network computing

differential privacy

NetDP: In-Network Differential Privacy for Large-Scale Data Processing

Abstract

Journal

Organization