Verifiable Dropout: Turning Randomness into a Verifiable Claim

2025-12-27

AI Summary

In cloud-based AI training, intrinsic nondeterminism from stochastic operations (e.g., dropout) enables adversaries to conceal malicious tampering within “natural randomness,” rendering existing logging-audit mechanisms incapable of verifying the honesty of random-value generation or application—while also failing to simultaneously preserve data and model privacy. Method: This paper introduces the first verifiable modeling of dropout randomness as cryptographic statements and proposes an embedded zero-knowledge auditing protocol within the training pipeline. Leveraging cryptographic hashing and a deterministic pseudorandom number generator (PRNG), it binds a trusted seed and generates succinct zero-knowledge proofs for post-hoc verification of stochastic operation execution. Contribution/Results: The scheme guarantees randomness unbiasness and consistency, incurs verification overhead under 0.8% of total training time, and leaks neither model parameters nor training data. It is the first solution enabling auditable stochastic operations in deep learning under strong privacy constraints.

Technology Category

• Machine Learning: Privacy
• Reasoning under Uncertainty: Stochastic Optimization
• Computer Vision: Adversarial Attacks & Robustness

Application Category

• Security and Privacy: Data transparency and provenance
• User Modeling, Personalization and Recommendation: Accountability, Transparency, and Ethics for personalization
• Responsible Web: Data and user privacy-enhancing technologies for the Web

Abstract

Modern cloud-based AI training relies on extensive telemetry and logs to ensure accountability. While these audit trails enable retrospective inspection, they struggle to address the inherent non-determinism of deep learning. Stochastic operations, such as dropout, create an ambiguity surface where attackers can mask malicious manipulations as natural random variance, granting them plausible deniability. Consequently, existing logging mechanisms cannot verify whether stochastic values were generated and applied honestly without exposing sensitive training data. To close this integrity gap, we introduce Verifiable Dropout, a privacy-preserving mechanism based on zero-knowledge proofs. We treat stochasticity not as an excuse but as a verifiable claim. Our approach binds dropout masks to a deterministic, cryptographically verifiable seed and proves the correct execution of the dropout operation. This design enables users to audit the integrity of stochastic training steps post-hoc, ensuring that randomness was neither biased nor cherry-picked, while strictly preserving the confidentiality of the model and data.

Problem

• Ensures dropout randomness is verifiable, not exploitable for deniability.
• Binds stochastic operations to cryptographic seeds for integrity verification.
• Enables privacy-preserving audit of training without exposing sensitive data.

Innovation

• Zero-knowledge proofs verify dropout mask integrity
• Cryptographic seeds bind stochastic operations deterministically
• Privacy-preserving audit of randomness without exposing data