FairLend: Privacy-preserving gender-fairness auditing for loan processing

Algorithmic credit-scoring systems may reproduce historical disparities while operating under strict borrower-data confidentiality requirements. Removing protected attributes such as gender from operational decision pipelines can reduce direct discriminatory use, but also makes group-disaggregated fairness auditing difficult. This paper presents FairLend , a privacy-preserving protected-attribute auditing framework for loan processing. FairLend separates the operational Loan Processing Unit (LPU) from an independent Fair Lending Auditor (FLA): the LPU processes authorised lending information without receiving gender in plaintext, while the FLA receives only encrypted aggregate group and outcome statistics. FairLend represents gender as a BFV-encrypted one-hot integer vector and aggregates protected-attribute statistics directly through ciphertext addition at multiplicative depth zero, yielding exact integer counts after decryption. Non-interactive zero-knowledge proofs (NIZKPs) support account-credential validity, credit-score authenticity, and application-component binding. A concrete Groth16 instantiation of the account-validity relation is implemented and benchmarked. Evaluation uses the LendingClub dataset with a controlled synthetic protected attribute correlated with non-sensitive proxy variables. Across the complete five- α 1 × ten-seed logistic-regression grid on the 177,489-record held-out test partition, all 600 encrypted aggregate-count comparisons exactly match their plaintext counterparts, producing zero demographic-parity and equalised-odds reconstruction error. Repeated-seed plaintext experiments additionally quantify proxy-inference and fairness sensitivity for logistic regression and random forest. FairLend is therefore a proof-of-concept privacy-preserving fairness-auditing architecture rather than a bias- mitigation mechanism or fully end-to-end encrypted credit-scoring system.