A team of researchers from the IMDEA Software Institute, Carlos III University of Madrid, and NEC Laboratories Europe has introduced a new framework that promises to improve the efficiency and practicality of verifiable computation. This work is detailed in the paper “Modular Sumcheck Proofs with Applications to Machine Learning and Image Processing,” which provides general proof systems and solutions tailored to specific applications in artificial intelligence and image processing. addresses the scalability and modularity challenges faced by both.
context
Verifiable computation consists of a set of cryptographic techniques that provide an unforgeable guarantee that a third party, such as a company or a cloud server, has performed the correct processing of user data. Proving that images or videos have been edited, that predictions made by artificial intelligence are based on audited models, or that only customer-provided data is used to determine creditworthiness , is an example of what these technologies enable. Additionally, verifiable computations are compatible with data privacy, so for example, the algorithms used by the server for the computations remain confidential.
Verifiable computation provides integrity, fairness, and privacy, which are essential properties for applications that outsource data processing tasks. Among the possible solutions are general proof systems, such as those used in some blockchains, but these have scalability issues when dealing with computations with large amounts of data. On the other hand, solutions designed specifically for these applications, while efficient, are often incompatible with each other, making them difficult to scale up or integrate into larger data processing chains.
the study
Researchers have introduced a new framework that aims to fill this gap by combining the performance benefits of custom solutions with the versatility of general-purpose test systems. At its core is a modular approach to verifiable computation of sequential operations based on a new cryptographic primitive known as Verifiable Evaluation Scheme (VE).
The researchers demonstrate the practical application of the framework in artificial intelligence by proposing a new VE that can adapt to convolution operations and handle multiple interconnected input/output channels. “Our protocol can be easily integrated into data processing chains, allowing for example the complete validation of predictions made by convolutional neural networks (CNNs), which are the basis of most artificial intelligence models,” he said. said David Balbás, PhD student at IMDEA. Software and researchers in this study. Furthermore, this paper also introduces a new His VE for image processing that can efficiently verify editing and retouching such as cropping, blurring, rescaling and other more complex operations.
The team created a prototype application for a test system that significantly improved existing technology. “In addition to introducing theoretical innovations in the algorithm, our benchmarks also demonstrate that proof generation is 5x faster and verification is 10x faster than the best existing solution to date. “Okay,'' explains Damien Robissout, Research Programmer at Institute IMDEA Software. Co-author of the study.
These results not only improve the efficiency and scalability of cryptographic proof systems, but also provide new solutions for ensuring the integrity, fairness, and privacy of data processing tasks in various applications of artificial intelligence and image processing. Open up possibilities. The application generated in the research is open source and its modular nature paves the way for extension and integration into different tools in the data processing chain. In this way, researchers pave the way for the versatile and robust deployment of verifiable computation in applications as diverse as financial ethics, personal data protection, and artificial intelligence regulation.
Disclaimer
This press release is the result of a collaboration between the ERC grant host institution and the European Research Council Executive Agency (ERCEA). The online project information and links contained in this press release are correct as of the date of publication. ERC is not responsible for outdated information or obsolete websites. Neither ERCEA nor its agents accept responsibility for the use of the information contained in this press release or for any errors that may remain in the text despite the care taken in its preparation.
