Case Studies

Blockchain Applications

Blockchain Based Privacy-preserving Computing

Figure 1: Universal secure computing over cloud, blockchain, and edge computing

The applications of MesaTEE stack can range from single-instance trusted computing to large scale distributed networks. In this sense, MesaTEE stack is the best fit for the computing backends of privacy-preserving blockchains -- as shown in Figure 1, wherever the data/code flow to, whenever the data/code get processed, and regardless of the underlying platform, the data/code owner can always be assured about the confidentiality and integrity protection.

Trusted Consensus with Extreme Performance

Figure 2: MesaTEE for blockchain services

Figure 2 demonstrates an example blockchain consensus system based on MesaTEE SGX (Level 5). Not only blockchain, other distributed computing scenarios requiring consensus or fault-tolerance can also leverage MesaTEE. Specifically, we utilize Raft over the mutually trusted TLS channels to establish consensus, and committed logs are written to each local database as the transactions are appended to the “chain”. In this case, the confidentiality and integrity of the records on chain are well protected from the platforms, and the remote attestation during the TLS communications guarantees that the consensus logic is as honest as expected. The overall blockchain can survive from data lost or unavailability as long as the majority of the participating nodes are still healthy (in Figure 2, the blockchain can tolerate one node’s failure). This is way more efficient and faster than other Proof-of-Work (PoW) based consensus solutions.

Sensitive Data Computing over Public Cloud

Figure 3: MesaTEE for public cloud services

Losing data confidentiality and integrity has been a major concern for public cloud users. MesaTEE can effectively solve this problem by establishing remotely-attestable trusted secure computing environments.

Figure 3 depicts MesaTEE’s deployment for public cloud services. Multiple TEE server nodes are grouped together to serve as a cloud cluster, which can be scheduled by orchestration framework like Kubernetes.

Secure Multi-party Computation

Figure 4: Traditional SMC compared to MesaTEE based SMC

When cross-department or cross-company data collaboration happens, privacy concerns arise. Thus secure multi-party computation (SMC) has become more and more important nowadays to enable joint big data analyses. However, traditional crypto-based SMC has quite a few limitations, and MesaTEE can solve them effectively, as shown in Figure 4:

1. once pre-computed for a given collaborative analyses, the confidential information can only be utilized by the targeted peers, and it’s difficult to add dynamic collaborators; however MesaTEE allows trusted and secure computing upon plaintext, thus does not impose such restriction.
2. once pre-computed for a given computation scheme, the confidential information can only perform determined algorithms, without the flexibility to add arbitrary follow-up computations; again MesaTEE’s trusted and secure computing upon plaintext does not impose such restriction.
3. once pre-computed for a given computation scheme, the confidential information can only be shared/computed by a fix number of participants, without the flexibility to add or remove participants; MesaTEE’s trusted and secure computing does not impose such restriction.
4. each of the participating parties needs to exchange information, which results in non-negligible latency and throughput downgrading; with MesaTEE, each data owner only needs to exchange information with the central trusted node, eliminating unnecessary communications.

Figure 5: MesaTEE’s superior performance on PSI, compared to traditional crypto-based approaches

Figure 5 uses Private Set Intersection (PSI) as the example to demonstrate MesaTEE’s great advantages over traditional SMC solutions. As a matter of fact, the more the participants, the larger the dataset to be analyzed, the better performance MesaTEE will obtain compared to traditional approaches.

Secure Spark Backend

Figure 6: MesaTEE’s integration as the Spark backend

As shown in Figure 6, we have also used Rust to re-write Spark backend executor in MesaTEE SGX (Level 5). Distributed big data analyses can thus enjoy strong confidentiality and integrity guarantees. This is extremely helpful when analyzing critical privacy dataset. We have also added oblivious protection to the framework, so that even if the platform owner observes the traffic flowing between the executors, he/she cannot learn any information from the pattern.

In order to achieve the same security promise, traditional crypto-based solutions usually cost days or even months to finish the workload that can be done in one second by the legacy Spark; but MesaTEE can finish execution in 4-20 seconds.

Secure Key Vault and HSM

Applications and services use cryptographic keys and secrets to help keep information secure. MesaTEE Key Vault can safeguard these keys and secrets, similar to hardware security modules (HSMs).

MesaTEE Key Vault can be conveniently built on all security levels 2-5, providing secret management (securely store and control accesses to tokens, passwords, certificates, API keys, and other secrets), key management (create and control encryption keys), and certificate management (provision and manage certificates).

Secure AI Computing

Figure 7: MesaTEE’s application in secure AI computing and the performance overhead is acceptable

We have already supported GBDT, Linear Regression, as well as neural networks in MesaTEE SGX (Level 5). Because we have ported TVM and Anakin in SGX, CNN/RNN models generated by popular AI frameworks (e.g. Tensorflow, Caffe, etc.) can be conveniently converted and loaded into SGX. As a result, both the model and the data can be well protected against the hostile environment. Although our major focus is on inference, we do support model training within SGX as well.

Performance is critical for AI applications, so we have also done a lot of optimizations to boost the efficiency, including model compression, model pipelining, etc. Figure 7 shows the time consumption of running ResNet 50 in MesaTEE SGX compared to running in the normal world (outside SGX) -- only 20% overhead is added. With this little addition, no other security solution ever can provide the same confidentiality and integrity promises.

For more details about MesaTEE trusted secure AI computation, one can take a look at our demo video: