Props for Machine-Learning Security

Ari Juels, Farinaz KoushanfararXiv 2024 · DOI · arXiv

Introduction

Machine learning (ML) is rapidly approaching a "data wall." While the surface web has provided a wealth of training data, we are reaching the limits of publicly accessible, high-quality information. To continue the trajectory of improvement, ML systems must tap into the "deep web"—private, authenticated data sources like electronic health records (EHRs) and financial statements. However, accessing this data presents a dual challenge: model developers need to verify data authenticity to prevent poisoning, while users require robust privacy guarantees.

The Props Framework

The authors introduce Props (protected pipelines), a framework designed to bridge this gap using privacy-preserving oracle technology. Props enable secure data pipelines that prove data provenance and integrity without requiring changes to existing web infrastructure.

The framework relies on two core technical pillars:

  1. Secure Data Sourcing: To ensure data entering the pipeline is authentic, Props utilize Privacy-Preserving Oracles. These can be implemented via hardware-based Trusted Execution Environments (TEEs) like Intel SGX or cryptographic methods like zkTLS (e.g., DECO). These tools allow a user to prove that a specific piece of data was received over a TLS connection from a trusted source (like a bank) without revealing the entire session.
  2. Pinned Models: For inference, Props use "pinned models" to provide a proof that a specific model $M$ was executed in a specific environment $E$ on the verified data $X$, producing output $Y$.

Applications and Use Cases

The Props architecture enables several transformative applications in the ML ecosystem, primarily focused on unlocking training data and enabling secure inference.

Unlocking Training Data

Props allow users to contribute private data to training sets while maintaining verifiable provenance. As shown in Figure 1, a user (Alice) can retrieve her medical records ($X$) from a provider (BigHospital) and submit them to an ML training system (MediModels). The Prop provides a proof that the data is authentic, protecting the trainer from fabricated data while allowing Alice to maintain control over her sensitive information.

Figure 1: Alice uses a Prop to provide verified data from BigHospital to an ML training system (MediModels) while maintaining provenance.

Secure Private Inference

Props also facilitate private inference, where a user receives a service based on their private data without exposing that data to the service provider. In Figure 2, Bob retrieves financial data ($X$) from BigBank. He runs a pinned model $M$ locally to generate a result $Y$ (e.g., a loan eligibility score). He then sends $Y$ along with a Prop proof to a third party (PrivaLoan), who can verify the result's validity without ever seeing Bob's raw bank statement.

Figure 2: Bob performs secure private inference, using a Prop to prove to PrivaLoan that his result $Y$ was correctly derived from authentic BigBank data.

Key Insights

  • Tapping the Deep Web: The deep web is estimated to be two orders of magnitude larger than the surface web. Props provide the technical "keys" to unlock this data for ML.
  • Integrity and Defense: By verifying data provenance, Props mitigate the risk of model poisoning from malicious users. Furthermore, they constrain adversarial attacks by requiring inputs to be verified, unmodified documents from trusted sources.
  • User Agency: Props support "contextual integrity," allowing users to apply filters (like redaction) while still proving the underlying data's authenticity.

Limitations and Future Directions

While promising, Props face several hurdles. Hardware-based TEEs are susceptible to side-channel attacks, and cryptographic zkTLS remains computationally expensive for large-scale applications. Additionally, while Props solve technical verification issues, legal questions regarding data ownership and regulations like GDPR remain complex. Future work must also address "model consistency" to handle rapidly updating models in production environments.

Conclusion

Props offer a path forward for the next generation of ML, enabling the use of high-quality private data while upholding the principles of integrity and privacy. By remaining infrastructure-independent, they provide a scalable solution for the looming data scarcity crisis.

Summary updated

Abstract

We propose protected pipelines or props for short, a new approach for authenticated, privacy-preserving access to deep-web data for machine learning (ML). By permitting secure use of vast sources of deep-web data, props address the systemic bottleneck of limited high-quality training data in ML development. Props also enable privacy-preserving and trustworthy forms of inference, allowing for safe use of sensitive data in ML applications. Props are practically realizable today by leveraging privacy-preserving oracle systems initially developed for blockchain applications.
BibTeX
@misc{juels2024props,
  author = {Ari Juels and Farinaz Koushanfar},
  title = {{Props for Machine-Learning Security}},
  year = {2024},
  eprint = {2410.20522},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2410.20522},
}
Figure