- Title
- Mitigation of Gradient Inversion Attacks in Federated Learning with Private Adaptive Optimization
- Creator
- Lewis, Cody; Varadharajan, Vijay; Noman, Nasimul; Tupakula, Uday; Li, Nan
- Relation
- 2024 IEEE 44th International Conference on Distributed Computing Systems (ICDCS). Proceedings of the 2024 IEEE 44th International Conference on Distributed Computing Systems (ICDCS) (Jersey City, NJ 23-26 July, 2024) p. 833-845
- Publisher Link
- http://dx.doi.org/10.1109/ICDCS60910.2024.00082
- Publisher
- Institute of Electrical and Electronics Engineers (IEEE)
- Resource Type
- conference paper
- Date
- 2024
- Description
- With the rapid advancements in machine learning as well as an increase in the awareness of privacy issues, federated learning has emerged to be an important paradigm as it greatly reduces the amount of data that need to be directly shared as part of the learning process. However, recently federated learning has been shown to be susceptible to gradient inversion attacks, where an adversary can compromise privacy by recreating the data that lead to a particular client's update. In this paper, we propose a new algorithm, SecAdam, to mitigate such emerging gradient inversion attacks and enable the clients to perform adaptive gradient based training in a federated setting while retaining client gradient privacy. We have given theoretical proofs for these properties as well as providing extensive practical experimental results, which we have carried out on five different datasets using two different neural network architectures. The results from these experiments demonstrate the effectiveness of our proposed algorithm. The code used to implement our algorithm, the different experiments and their analysis are available at https://github.com/codymlewis/SecOpt.
- Subject
- federated learning; privacy; adaptive optimization; secure aggregation
- Identifier
- http://hdl.handle.net/1959.13/1518695
- Identifier
- uon:57349
- Identifier
- ISBN:9798350386066
- Language
- eng
- Reviewed
- Hits: 47
- Visitors: 48
- Downloads: 0
| Thumbnail | File | Description | Size | Format |
|---|