AI Bias Verification via Blockchain Hashing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning AI systems lack transparency, making it impossible for humans to understand how they arrive at predictions, and they may inadvertently use impermissible cohorts as factors in decision-making, leading to biased outcomes.
Innovation Solution
A method and system that utilize primary and supervisory machine learning models to verify that predictions are not based on biased cohorts by stripping biased data, hashing relevant information, and storing it in a blockchain to ensure compliance and transparency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning models are used to make predictions, then prediction accuracy is improved, but transparency and understandability of the decision-making process deteriorates
Solution Approach 1:
The patent introduces an intermediary verification system that includes: (1) a hash utility that creates cryptographic hashes of the deep learning model's input data and parameters, (2) a blockchain that stores these hashes immutably, and (3) a verification mechanism that compares hashes to detect whether biased data influenced the prediction. This intermediary layer maintains the deep learning model's predictive accuracy while providing transparent, verifiable evidence about the data processing journey.
Solution Approach 2:
The patent implements a feedback mechanism where the verification system continuously monitors and records the deep learning model's processing of data. By hashing input data, model parameters, and intermediate results, and storing them in a blockchain, the system creates a feedback loop that provides real-time transparency about whether biased cohorts influenced predictions, without interfering with the model's core predictive function.
2Measurement precision
If biased data sets are included in training, then model performance on specific cohorts is improved, but fairness and lack of bias deteriorates
Solution Approach 1:
The patent applies the extraction principle by removing biased data elements from the training and prediction process. The verification system identifies and extracts problematic cohort markers by hashing and comparing input data against known biased data sets. When biased data is detected, the system can exclude or weight-adjust these elements, allowing the model to maintain performance on legitimate predictive features while eliminating harmful bias against protected cohorts.
Solution Approach 2:
The patent implements preliminary anti-action by proactively detecting and preventing bias before it affects predictions. The verification system pre-identifies biased cohort markers in the training data and prediction inputs, creates cryptographic hashes of these identified biases, and stores them in the blockchain for comparison. This preliminary detection and marking of biased data prevents the bias from influencing model performance, allowing fair treatment of all cohorts while preserving legitimate predictive accuracy.
3Object-generated harmful factors
If verification mechanisms are added to detect bias, then fairness and transparency are improved, but system complexity increases
Solution Approach 1:
The patent replaces complex manual verification mechanisms with automated cryptographic hashing and blockchain-based validation. Instead of requiring intricate analysis of whether biased cohorts influenced predictions, the system uses hash functions to create unique digital fingerprints of input data, model parameters, and processing steps. These hashes are automatically compared against stored references, and the blockchain's inherent security properties provide verification without requiring complex analytical systems.
Solution Approach 2:
The patent changes the verification approach from analyzing complex data patterns to comparing cryptographic hash parameters. By transforming the verification problem into a parameter-comparison task (comparing hash values rather than analyzing data content), the system achieves bias detection with simpler, more efficient operations. The blockchain stores and manages these hash parameters, providing a straightforward verification mechanism that reduces overall system complexity.
Data Source
AI summary
A method including receiving data including an unknown vector including a data structure populated with unknown features describing a first user and a score predicted by a MLM trained using a prediction data set. The score represents a prediction regarding the first user. The prediction data set includes the unknown vector stripped of a biased data set. The data also includes a prediction whether the first user belongs to the cohort. The method also includes hashing information types used by the primary MLM and the supervisory MLM to produce a first hashed data, the information types including at least the unknown vector, the score, and the prediction. The method also includes combining the first hash and a schema to produce a compliance document. The method also includes hashing the compliance document to produce a second hashed data. The method also includes storing the second hashed data in a blockchain.