As there is an increasing need for efficient models to mine data of late given growing privacy concerns as well as growth of data velocity, veracity, variety and volume (Brisimi et al., 2018). Federated learning offers a decentralized computational framework which manages the available data as a part of a virtual (federated) database, avoids centralized data collection, data processing, and raw data exchanges, and has a strong predictive power (Brisimi et al., 2018). This paper looks into how Google can apply the federated learning model in order to improve the issue of user privacy.
The privacy issue is resolved by the Federated Learning model because it works without storing data of users in the cloud. When one thinks of standard machine learning, its approaches demand centralization of the training data on a single machine or, alternatively, in a specific datacenter. Google’s cloud infrastructures focused on data processing are some of the most robust and secure. At the same time, for those models which are trained from interaction with various mobile devices, they have introduced another method, which is Federated Learning. It enables mobile phones to learn in a collaborative manner a shared prediction model while it ensures that all training data are kept on device (McMahan & Ramage, 2017). It decouples the ability to carry out machine learning from the necessity to store the data in a single cloud.
Use your promo and get a custom paper on
"Google AI at Its Best".
How does it work? This essentially works beyond local models’ use as they make predictions on phones or other mobile devices as model training is brought to these devices as well. It occurs in the following manner: one’s mobile device downloads the up-to-date model, then improves it through learning from the available data on this device, and after this sumps up the changes. The latter comes as a focused small update. It is only this very model update that is sent to the cloud with help of encrypted communication, where it is instantly averaged with a set of other updates from users and, in this way, makes the shared model better (McMahan & Ramage, 2017). The benefit here is that all the data involved in training remains on one’s device and the cloud does store any individual updates. It is, however, worth noring that the system in this framework is obliged to communicate as well as aggregate all updates to models in fault-tolerant, secure, scalable, and efficient way.
Moreover, privacy gets enhanced owing to the use of a Secure Aggregation Protocol. This procol utilizes cryptographic techniques so that the server tasked with coordination can decrypt on average if 100 or 1000 users have taken part. Essentially, since the model focuses on the use by mobile phones, the benefit of introducing the Secure Aggregation Protocol is as follows: no individual’s mobile phone’s update is inspected prior to averaging (McMahan & Ramage, 2017). This enables effective work with deep-network-sized problems and also allows working in the circumstances of real-world constraints of connectivity. Google engineers designed Federated Averaging in the way that only an average update is required by the coordinating server, which makes it possible to use Secure Aggregation.
In conclusion, Federated Learning model is capable of solving the issue of privacy because of its inherently decentralized framework, which only yields relatively few features to classifiers, and allows efficient information processing with smartphones. In terms of privacy, Federated Learning is more practical because it doe snot rely on a central data repository and, in case with Google, uses a Secure Aggregation Proocol. It looks like this decentralized framework allows multiple holders of data to collaborate without explicit exchanges of raw data, which ensures its security and user privacy.
- Brisimi, T., Chen, R., Mela, T., Olshevsky, A., Paschalidis, I., & Shi, W. (2018). Fedeerated learning of predictive Electronic Health Records. International Jounral of Medical Informatics, 112, 59-67.
- McMahan, B. & Ramage, D. (2017). Federated learning: Collaborative machine learning without centralized training data. Google AI Blog. Retrieved from https://ai.googleblog.com/2017/04/federated-learning-collaborative.html.
