Image credit: Consilient
Machine learning underpins much of the artificial intelligence used today. It is fast, accurate, and exponentially improves the processes for which it is designed.
AI models are trained using data that is aggregated from multiple edge devices and then sent to a centralized server. The data is applied to machine learning algorithms to train the model, which updates predictions based on the new information.
That works well in a closed network scenario, but the prevalence of cloud computing has a major impact on machine learning. Cloud-based AI solutions can undercut user privacy. Using traditional machine learning methods, sensitive user data is sent to the cloud along with the training data.
To combat that problem, AI companies are turning to federated learning, a decentralized form of machine learning. Federated learning pulls together on-device AI, blockchain, and edge computing with Internet of Things technology. It can collect data from distinct locations, sometimes even different organizations, but it only shares the relevant training information to protect user privacy and other sensitive data.
Image credit: Consilient
Also known as collaborative learning, federated learning trains a central model using decentralized data. A local copy of the centralized machine learning application resides on the edge devices, and the model is trained locally doing the processing at the edge, not the centralized server. The training results are transferred through the cloud to the central server, but the sensitive user data is encrypted and inaccessible.
The centralized server aggregates the results from all the edge devices and updates its machine learning models. A copy of the newer model is shared with the edge devices and the process continues, further refining the machine learning model. Because it aggregates training data from multiple sources it can identify common patterns and build a robust model quickly while preserving user privacy
The user data is encrypted locally, and the encryption key is not on the centralized cloud server. The server is only able to decrypt the aggregated results. Because it protects sensitive data, federated learning is well-suited for the defense, telecommunications, financial, and pharmaceutical industries.
The use of federated learning can, in fact, improve artificial intelligence because it expands the data sources that train the models. For example, Intel® partner Consilient has developed Dozer, a federated learning solution that helps financial institutions detect fraud and money laundering.
Many of the current fraud detection systems are focused on the regulatory side, and they often still involve manual procedures. Even with those processes in place, McKinsey estimates that illegal transactions total $800 billion to $2 trillion annually, and authorities discover less than 1 percent of those illegal transactions.
Consilient’s Dozer, with Intel’s Software Guard Extensions (SGX), aggregates anti-fraud and anti-money laundering analytics from multiple financial institutions. Dozer improves the AI training models in the central database, while protecting user and bank data. Traditionally that information is siloed within each banking institution, but Dozer provides banks with analytical insights from across the entire sector. That enables earlier detection of bad actors or fraudulent schemes.
Similar benefits can be gained on the manufacturing floor, where computer vision is used to perform product inspections. Not only can manufacturers discover flaws more quickly, edge processing also prevents thousands of images from being transported to the cloud, thus improving latency. That saves manufacturers time, money, and network resources, while still improving the end product.