FedOD: Federated Outlier Detection via Neural Approximation

Yue Zhao, Sidharth Sharma, Zhihao Jia

Primary Area: general machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: outlier detection, federated learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: A novel machine learning system to support federated learning on non-neural-network outlier detection algorithms.

Abstract: Outlier detection (OD) is a crucial machine learning task with key applications in various sectors such as security, finance, and healthcare. Preserving data privacy has been increasingly important for OD due to the sensitivity of the data involved. While federated learning (FL) offers the potential in protecting data privacy, it is not yet available for most classical OD algorithms, such as those based on distance and density estimation. To address this, we introduce FedOD, the first FL-based system designed for general OD algorithms. FedOD effectively overcomes the privacy and efficiency challenges inherent in classical OD algorithms by automatically decomposing these algorithms into a set of basic operators and approximating their behaviors using neural networks. Given the inherent compatibility of neural networks with FL, the approximated OD algorithms also become capable of privacy-preserving learning without data exchange. With this design, FedOD supports over 20 popular classical OD algorithms and is readily extendable to other fields like classification. Evaluation on more than 30 benchmark and synthetic datasets demonstrates FedOD's accuracy and efficacy over state-of-the-art baselines---compared to existing OD systems, FedOD achieves up to 11x reduction in errors and 10x improvement in performance.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 8709