Abhishek Singh, Split Inference - Metrics, Benchmarks and Algorithms, ECCV'24

Abstract

Split Inference has emerged as a practical approach to split the computation across multiple parties for privacy and efficiency reasons. In this work, we tackle the question of how to evaluate data leakage when the output (activations/embedding) of a model is shared with untrusted parties.  In this work, we characterize different obfuscation techniques and design new attack models. We propose multiple reconstruction techniques based on distinct background knowledge of the adversary. We develop a modular platform that integrates different obfuscation techniques, reconstruction algorithms, and evaluation metrics under a common framework. Using our framework, we benchmark various obfuscation and reconstruction techniques for evaluating their privacy-utility trade-off. Finally, we release a dataset of obfuscated representations to foster research in this area. We have open-sourced code, datasets, hyper-parameters, and trained models that can be found at https://github.com/aidecentralized/InferenceBenchmark.