Data privacy regulations such as GDPR, CCPA, and HIPAA present a challenge for training AI systems on sensitive data, such as financial transactions, patient medical records and user device logs. Historical data is what “teaches” AI systems to identify patterns and make predictions, but there are technical barriers to using it without compromising a person’s identity.
A workaround that has become popular in recent years is federated learning. The technique forms a system on multiple devices or servers containing data without ever exchanging it, allowing collaborators to create a common system without sharing data. Intel recently partnered with Penn Medicine to develop a brain tumor classification system using federated learning, while a group of big pharma, including Novartis and Merck, built a federated learning platform to accelerate drug discovery.
Tech giants including Nvidia (via Clara) offer Federated Learning as a Service. But a new startup, DynamoFL, hopes to take on incumbents with a federated learning platform that focuses on performance, seemingly without sacrificing privacy.
“DynamoFL was founded by two PhDs from the Department of Electrical Engineering and Computer Science at MIT, Christian Lau and myself, who has spent the last five years working on privacy-preserving machine learning and hardware for machine learning,” CEO Vaikkunth Mugunthan told TechCrunch in an email interview. “We discovered a huge market for federated learning after receiving repeated job offers from large financial and technology companies trying to develop federated learning internally in light of emerging privacy regulations such as GDPR and CCPA . During this process, it was clear that these organizations were struggling to implement federated learning internally and we built DynamoFL to fill this gap in the market.
DynamoFL — which claims to have key customers in the automotive, IoT and finance industries — is in the early stages of its go-to-market strategy. (The startup currently has four employees and plans to hire 10 by the end of the year.) combat attacks and vulnerabilities in federated learning — such as “member inference” attacks that allow detection of data used to train a system.
Picture credits: DynamoFL
“Our personalized federated learning technology…allows[s] machine learning teams to tune their models to improve performance on individual cohorts. This gives C-suite executives greater confidence when deploying machine learning models that were previously considered black box solutions,” said Mugunthan. “This [also] differentiates us from competitors like Devron, Rhino Health, Owkin, NimbleEdge, and FedML who struggle with the common challenges of traditional federated learning.
DynamoFL also advertises that its platform is cost-effective compared to other privacy-preserving AI dot solutions. SSince federated learning does not require the massive collection of data on a central server, DynamoFL can reduce data transfer and computation costs, Mugunthan claims, for example by allowing a client to send only small files. incremental rather than petabytes of raw data. As an added benefit, this can reduce the risk of data leaks by eliminating the need to store large volumes of data on a single server.
“Common privacy-enhancing technologies, such as differential privacy and federated learning, have suffered from a perennial “privacy versus performance” trade-off, where the use of more robust privacy-preserving techniques during training of the model inevitably leads to a lower accuracy of the model. This critical bottleneck challenge has prevented many machine learning teams from adopting privacy-preserving machine learning technologies that are needed to protect user privacy while complying with regulatory frameworks,” Mugunthan said. . “DynamoFL’s custom federated learning solution addresses a critical barrier to machine learning adoption.”
Recently, DynamoFL closed a small seed round ($4.15 million at a valuation of $35 million) involving Y Combinator, Global Founders Capital and Basis Set; the startup is part of Y Combinator’s winter 2022 batch. Mugunthan says the proceeds will primarily go towards hiring product managers who can integrate DynamoFL’s technologies into future user-friendly products.
“The pandemic has highlighted the importance of quickly leveraging diverse data for emerging health crises. In particular, the pandemic has underscored how critical medical data must be made more accessible in times of crisis, while protecting patient privacy,” Mugunthan continued. “We are well positioned to weather the slowdown in technology. We currently have three to four years of track, and the tThis slowdown has actually helped our hiring efforts. The biggest tech companies were hiring the majority of Federated Learning’s top scientists, so the slowdown in big tech hiring presented us with an opportunity to hire top Federated Learning and IT talent. machine learning.
