Privacy and Applied Cryptography

Ensuring provable privacy can only be done by cryptographic protocols, but it is often hindered by significant computational overhead. While recent years have witnessed drastic improvements in runtime for provable and cryptographically secure privacy-preserving computing (P^3C), maintaining acceptable performance in practical applications remains a challenge. To enable real-world applications, there is a dire need to bridge the gap between P^3C and plaintext computation while defining the boundaries of protocol applicability. Although statistical methods can anonymize data with more manageable overhead, they offer limited privacy guarantees and remain vulnerable to sophisticated attacks.

ACES Lab is focused on building state-of-the-art privacy-preserving systems at the intersection of multiple cutting-edge domains, reaching the efficiency of statistical methods while maintaining the rigor of provable privacy. Our most prominent works utilize techniques such as Multi-Party Computation (MPC) and Fully Homomorphic Encryption (FHE) to ensure practical, end-to-end security.

Our recent research has primarily pivoted toward Zero-Knowledge Proofs (ZKPs). ZKPs allow a prover (P) to convince a verifier (V) that a computation f on a private witness w is correct without revealing any information about w. While the potential of ZKPs is limitless, their widespread adoption has been largely relegated to the blockchain. Our work aims to move ZKPs beyond the ledger by securing complex learning paradigms and real-world AI applications. To achieve this, we have focused on three separate research pillars:

1. HW/SW co-design: We conduct research on custom hardware acceleration. This includes the development of specialized accelerators designed to handle ZK friendly hash functions (e.g. Reinforced Concrete, Griffin, Rescue-Prime), significantly reducing the bottleneck of proof generation.

2. Verifiable Learning: We design cryptographically secure and verifiably robust training algorithms for several emerging learning paradigms, such as federated learning and split learning.

3. Model integrity & ownership: We utilize ZKPs to establish ownership of neural networks and the integrity of their outputs, providing cryptographically secure and robust watermarks that allow creators to protect IP and prove model provenance in the age of open source AI.

Our current thrust lies in investigating methods to push the boundaries of scalable ZKPs for machine learning. More specifically, we are exploring different methods of model manipulation techniques to partition massive LLMs across distributed environments, such as in compute centers. By pursuing parallelized sharding strategies, we aim to drastically decrease proof generation times for the next generation of massive scale verifiable AI.