At Q2BSTUDIO, a leading company in technology development and services, we specialize in creating innovative solutions for data processing and analysis. In this context, we explore advanced strategies for DP-based Fair Supervised Learning. This approach is key to ensuring that artificial intelligence and machine learning models are fair and do not introduce unwanted biases in their predictions.
In our experiments, we worked with various benchmark datasets in the machine learning community: COMPAS, Adult, and CelebA. These datasets contain sensitive attributes such as race and gender, allowing us to evaluate the fairness of the trained models.
For training, different differential privacy-based learning methods were used, including fair classification algorithms based on density estimation, mutual information, and maximum correlation. The implemented models included logistic regression for linear classification and deep neural networks to capture more complex patterns in the data.
The neural network architecture used varied according to the dataset:
1. For COMPAS, a multilayer perceptron (MLP) with 4 hidden layers and 64 neurons per layer was implemented.
2. In the case of the Adult dataset, an MLP with 4 hidden layers and 512 neurons per layer was used.
3. For CelebA, whose data format is based on images, ResNet-18 was used, an architecture optimized for this type of input.
The results obtained demonstrate that the use of fair supervised learning techniques based on differential privacy improves the fairness of models without compromising their performance. At Q2BSTUDIO, we remain committed to developing advanced technological solutions that prioritize transparency, fairness, and privacy in data processing. Our team continuously works to improve and apply new methodologies to ensure fair and efficient models in various industrial applications.
