Search papers, labs, and topics across Lattice.
This paper introduces a novel method for cryptographically committing all data entering and leaving AI data centers, utilizing network taps to compute hashes on information-carrying wires. The proposed architecture, termed the "Secure Gateway Device," effectively mitigates covert data exfiltration by addressing vulnerabilities such as analogue and timing side-channels without relying on mutually trusted processors. Key results indicate that this approach not only enhances compliance verification but also significantly reduces the risk of unauthorized data leakage during AI workload processing.
Covertly exfiltrating AI workload results is rendered infeasible, thanks to a new architecture that eliminates the need for mutually trusted processors.
In preparation for potential international agreements on artificial intelligence, the development of verification infrastructure for AI data centres is vital. We propose a method for cryptographically committing all information entering and leaving a data centre: Hashes are computed by network taps placed on all the information-carrying wires between the cluster and the outside world, enabling an auditor to retroactively challenge the preimage data to be sent to a privacy-preserving verification facility performing compliance checks. Our goal is to make it infeasible to covertly exfiltrate the results of undisclosed workloads in the cluster through the tapped wires. To this end, we specify the architecture of a ``Secure Gateway Device'', which handles the erasure of covert channels that post-hoc verification on hashed data cannot address: analogue and timing side-channels, as well as steganography in network protocol headers. The architecture eliminates the need for any processors trusted by both the Prover and the Verifier, leveraging passive optical fibre splitters and coin-flip protocols for random number generation where needed. We expect development costs of a demonstration device to be roughly equivalent to the cost of a small team of engineers for a few months, with a comparatively small bill of materials.
