Tamper-Proof Privacy Auditing for Artificial Intelligence Systems

Andrew Sutton, Reza SamaviIJCAI 2018 · DOI

TL;DR

This paper introduces a novel method for creating tamper-proof privacy audit logs for AI systems using blockchain technology. By leveraging Linked Data and Bitcoin's blockchain, they enable auditors to verify the integrity and authenticity of audit logs, ensuring accountability in data-sharing environments.

What Problem Does This Paper Address?

In collaborative healthcare and AI research, vast amounts of sensitive personal data are shared. Data Sharing Agreements (DSAs) govern this sharing, and privacy audit logs are crucial for monitoring compliance. However, a significant trust issue arises when auditors might collude with participants (researchers or even the AI systems themselves) to alter or hide non-compliant actions within these logs. This collusion, or even inadvertent breaches, can undermine the entire auditing process, making it difficult to hold parties accountable for privacy violations. The paper addresses the need for a mechanism that guarantees the integrity and non-repudiation of these audit logs, ensuring they accurately reflect all actions.

Methods

The proposed solution builds upon the L2TAP (Linked Data-based Privacy Audit Log) framework, which already uses Linked Data principles to represent privacy events. The core innovation lies in integrating blockchain technology to secure these logs.

Here's a breakdown of their approach:

  • Logger Signing: Each privacy log event generated by the logger is digitally signed using the logger's private key. This ensures non-repudiation, meaning the logger cannot later deny having generated the event.
  • Signature Graph Generation: A "signature graph" is created, containing assertions about the event's signature, including the signature value and verification methods. This graph is then signed by the logger.
  • Integrity Proof Generation: A cryptographic hash (integrity proof digest) of the event graph and its corresponding signature graph is computed. This digest acts as a unique fingerprint for the event.
  • Blockchain Integration: These integrity proof digests are then written to the Bitcoin blockchain through transactions. The OP_RETURN opcode in Bitcoin transactions is used to store this data.
  • Block Graph Generation: After a transaction is confirmed on the blockchain, the data is queried, parsed, and stored as a "block graph." This block graph contains the integrity proof digest and information about the block it's stored in. These block graphs are stored in a quad store for easy auditor access.
  • Log Integrity Verification: Auditors can then query the quad store for block graphs. By recomputing the integrity proof digest of a log event and comparing it to the stored digest on the blockchain, they can verify if the log event has been tampered with. If the digests match, the event's integrity is confirmed. The signature verification is also performed to ensure authenticity.

Datasets

Not discussed in the paper. The authors mention generating synthetic L2TAP logs for their experiments, which consist of events like log initialization, privacy preferences, access requests, and actual access.

Key Insights and Ideas

  • Blockchain as an Immutable Ledger for Audit Trails: The paper effectively demonstrates how blockchain's inherent immutability can be leveraged to create a trustworthy audit trail, solving the problem of log tampering.
  • Bridging Linked Data and Blockchain: The integration of Linked Data (for rich event representation) with blockchain (for tamper-proofing) is a key contribution, creating a robust system for privacy auditing.
  • Decentralized Trust for Auditors: By storing integrity proofs on a public blockchain, the system reduces reliance on a single, potentially compromised, central authority for trust. Auditors can independently verify log integrity.
  • Scalability through Block Graphs: Storing integrity proofs on the blockchain and then creating block graphs in a local quad store significantly reduces the burden on auditors, preventing them from having to search the entire blockchain for individual proofs.

Key Results

The experimental evaluation focused on the scalability of the log integrity verification procedure.

  • Linear Time Growth: The experiments demonstrated a linear growth in execution time for verifying the integrity and signature of log events as the number of events increased. This suggests the approach is scalable.
  • Worst-Case Scenario: The authors opted to demonstrate a worst-case scenario by verifying the entire log rather than a subset, which still yielded promising scalability results.
  • Specifics: While exact numbers for verification times are presented in Figure 4, the key takeaway is the linear relationship between the number of events and the time taken for verification. For example, verifying 10,000 events took approximately 1500 seconds.

Limitations and Future Work

  • Bitcoin Limitations: The authors acknowledge Bitcoin's limitations in terms of cost, speed, and scalability. They used Bitcoin primarily to demonstrate feasibility.
  • Optimized Implementations: Future work includes comparing other blockchain technologies for better transaction fees, scalability, and smart contract support.
  • Merkle Trees: For very large logs, Merkle trees are suggested to reduce storage and transaction requirements on the blockchain, though this would increase the effort for log integrity verification.
  • Formalizing Trade-offs: Investigating the trade-offs between using hash trees and the verification effort is identified as an interesting optimization problem.
  • Internal Attacks: While the system protects against external tampering, the paper notes that an internal attacker could still attempt to subvert the verification process by generating and signing fake log events, storing them, and then generating corresponding blockchain entries. This would require a sophisticated internal attack.

Summary updated

BibTeX
@inproceedings{sutton2018tamper,
  author = {Andrew Sutton and Reza Samavi},
  title = {{Tamper-Proof Privacy Auditing for Artificial Intelligence Systems}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/756},
}
Figure