Abstract
Abstract This paper presents the Auditable Zero-knowledge Transformer (AZT) framework for privacy-preserving and auditable tax fraud detection. AZT combines transformer-based anomaly detection with zero-knowledge proof (ZKP) verification so that a tax authority or regulator can verify fraud-detection outcomes without accessing sensitive taxpayer records or proprietary model parameters. The framework is scalable in the specific sense of low-latency audit verification: proof verification remains sub-second, whereas proof generation is intentionally performed asynchronously after local inference. Model integrity is enforced through Merkle-root commitments to authority-approved parameters, and the ZKP statement proves that the committed transformer was executed correctly and that the resulting risk score satisfies a public audit threshold. Experiments on UCI-TFD, IRS-Pub, and CorpPay compare AZT with classical machine-learning baselines, including Random Forest and XGBoost, and with an equivalent plaintext transformer. Detection quality improves by up to 5.3% in F1-score over classical machine-learning baselines, while the circuit-compatible AZT inference incurs only about 0.5% F1-score degradation relative to the plaintext transformer baseline. Overall, this work advances secure AI for digital governance by integrating modern deep learning with cryptographic verification, offering a practical foundation for fraud-detection systems in which transparency and confidentiality must be satisfied simultaneously.