Last week, Toon Segers (Roseman Labs' Co-Founder and Head of Customer Success) attended the academic conference TPMPC 2024 (Theory and Practice of Multi-Party Computation) alongside other Roseman Labs colleagues. Hosted in Darmstadt by TU Darmstadt’s ENCRYPTO group, the conference was a goldmine of insights in the fast-advancing field of MPC. Here are Toon's key takeaways:
The TPMPC conference proved that complex tasks like secure inference with Large Language Models (LLMs) are feasible with today's hardware. This was particularly highlighted in the talk titled "Sigma: Secure GPT Inference with Function Secret Sharing" by Neha Jawalkar, from the Indian Institute of Science, and colleagues from Microsoft Research, who showed that running inference operations on an encrypted 13 billion parameter model is feasible (using MPC servers with GPUs) with inference times of a few seconds per token! You can read their paper here.
This is a big achievement that builds on a large body of work in research on AI on encrypted data. It also shows that MPC is ready for cost-effective enterprise workloads, and is getting ready for modern AI on encrypted data in the coming years.
Accelerations in the MPC domain
During the conference, we saw accelerations in the domain of convolutional neural networks, cloud deployments, and structure-aware private-set intersection, or set operations that are optimized for specific data-structures, such as hierarchical relationships, graphs and ordered data.
Our colleague Meilof Veeningen contributed a talk about GDPR-compliant processing of mobility data using secure joins and groupby-operations in a convenient pandas/scikitlearn-like analysis environment that data scientists working with Python know and love. This is to name just a few topics that are relevant for broadening the scope of MPC applicability.
Innovations by Roseman Labs
The field has a bright future, which is accompanied by recent innovations and acceleration from our own Roseman Labs team in the MPC domain. As mentioned, these include scalable tabular operations, regular expressions on large amounts of text data and acceleration of traditional machine learning algorithms, such as regression models, K-Nearest Neighbours, tree-based models, and neural nets – for both training and inference on encrypted data. The beauty of these building blocks is that we can safeguard both the training data and the query, as well as the AI-model, depending on the sensitive nature of the use-case and the performance requirements.
Ready for today's enterprise workloads and tomorrow’s challenges
The above highlights that MPC is getting ready for today and tomorrow’s enterprise workloads. Particularly those that are SQL-like or AI-driven. We expect that very soon, MPC will be able to ingest more data, and more types of data such as images, free text, and DNA sequences, and process these at higher speeds. Either through novel cryptographic protocols, parallelism and (commodity) hardware-acceleration. In short: The field of MPC is gearing up to meet these challenges head-on.
Closing remarks
TPMPC confirmed that MPC has an extremely bright future and, after decades of building the foundation, is poised to deliver on its promises. At the same time, the topic of privacy-centric AI is front and centre. For instance, with Apple’s announcement of Private Cloud Compute - its effort to provide a secure environment for running sensitive AI-models. Against this fast-moving background, companies like Roseman Labs are committed to delivering a scalable encrypted computing experience.
In conclusion, the TPMPC 2024 conference was a testament to the exciting future of MPC. As we continue to make strides in this field, we can look forward to more secure, efficient, and versatile computing solutions.
Generate new insights on sensitive data with Roseman Labs’ secure Multi-Party Computation technology. Want to find out how your organization can do that? Contact us using the form below.