Establishing Reputations through Prototypes: Impartiality and Robustness in the Federated Context

• Grant number: 24/22560-2
• Support Opportunities: Scholarships abroad - Research Internship - Scientific Initiation
• Start date: March 24, 2025
• End date: July 23, 2025
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: André Carlos Ponce de Leon Ferreira de Carvalho
• Grantee: Arthur Hiratsuka Rezende
• Supervisor: Carlos Manuel Milheiro de Oliveira Pinto Soares
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• Institution abroad: Universidade do Porto (UP), Portugal
• Associated to the scholarship: 24/08236-8 - Intelligent Architecture for Water Distribution: Robust Clustering to Demand Variations under Federated Learning, BP.IC

Abstract

A holistic perspective is vital in the search for robust Artificial Intelligence (AI) solutions. In the context of Smart Cities, AI implementations must adapt to variations in urban scenarios. Real consumption data from the city of São Paulo, obtained through a partnership with Sabesp, reveals discrepancies between the company's consumption estimation model and actual data. Exploratory analysis identifies the emergence of a new class of properties (apartments up to 50m²) resulting from the 2014 revision of São Paulo's "Plano Diretor". This development has led to the degradation of the estimation model over time, providing a clear example of Concept Drift, a phenomenon to be studied with the foreign research group.Furthermore, there is a strong correlation between water consumption and the construction standard of properties (derived from IPTU tax data), which can be directly linked to the social strata of residents. High-end properties are found to consume up to three times more water than modest residences. Since the ongoing project in Brazil aims to detect anomalies in the water network (such as leaks), removing biases from the data is critical. For example, if the models are biased toward higher-standard properties (which consume more), failures in detecting issues in lower-income areas (with less available data and lower consumption) may occur. In such cases, an increase in water volume due to a leak might be misclassified as normal consumption, resulting in pressure drops and reduced service quality.This proposal addresses concerns about providing a system unbiased toward specific groups (Fairness) and about the degradation of models over time (Concept Drift). A thorough investigation of sources of bias and heterogeneity among participants in Federated Learning (FL)-topics intrinsically related to Fairness and Concept Drift-is conducted. It's found in the literature the so called prototypes as a strategy to address heterogeneity and domain shifts among participants in FL, yielding promising results in model robustness by mitigating data biases. It is hypothesized that the adoption of unbiased prototypes can enhance resilience to domain shifts. Furthermore, the research hypothesizes that prototypes could be useful in detecting Concept Drift (via a proposed reputation policy) and promoting Fairness in the federated context.The goal of the research abroad is to assess the benefits and challenges of prototyping within the Intelligent Architecture for water distribution currently under development in Brazil. The aim is for this strategy to enhance the robustness of the ongoing project by detecting predictive model degradation (as observed in the Sabesp regressor), with significant impacts on the project's applicability in real-world scenarios. The contribution of the foreign research group will be crucial for achieving these objectives.