Scholarlink

Crime rates in India have surged rapidly in recent years. This increase has significantly strained the judicial and investigative systems, resulting in numerous cases remaining unresolved or delayed. Prolonged case hearings have a profound impact on the nation’s social and judicial institutions, disproportionately affecting vulnerable communities that may lack the resources to pursue lengthy legal battles, thereby undermining their faith in the justice system. First Information Reports (FIRs) serve as critical records of criminal incidents, containing valuable information about suspects and case details. Analyzing FIR data can be instrumental in predicting potential convictions. However, due to data privacy concerns, police stations do not disclose or provide open access to FIR data. This necessitates the development of methods that analyze FIR data while preserving privacy, leading to the adoption of Federated Learning (FL) as an effective solution. FL allows data to remain decentralized while enabling collaborative model training, thus addressing privacy challenges. In this study, historical FIR data from Karnataka State is utilized to train Machine Learning (ML) and Deep Learning (DL) models for predicting convictions. Various classification algorithms are implemented, including Logistic Regression, Decision Tree, Naïve Bayes, Random Forest, XGBoost, and Neural Networks. Random Forest achieved the highest classification performance, with an accuracy of 98% on test set, highlighting the model’s potential for deployment in privacy-sensitive environments. On implementing various FL methods on Random Forest, 95% was achieved with FedAvg. This study demonstrates the feasibility of privacy-preserving conviction prediction using decentralized FIR data.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Federated Learning Approach for Predicting Conviction Using FIR Data

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Aswin Ram Kumar Gannina,
  • Chaitra Sai Chakravarthi Ganapaneni,
  • Rishi Deep Ghanukota,
  • Venkata Sasidhar Kommu,
  • Uma Mahesh Palla,
  • P. Malathi

摘要

Crime rates in India have surged rapidly in recent years. This increase has significantly strained the judicial and investigative systems, resulting in numerous cases remaining unresolved or delayed. Prolonged case hearings have a profound impact on the nation’s social and judicial institutions, disproportionately affecting vulnerable communities that may lack the resources to pursue lengthy legal battles, thereby undermining their faith in the justice system. First Information Reports (FIRs) serve as critical records of criminal incidents, containing valuable information about suspects and case details. Analyzing FIR data can be instrumental in predicting potential convictions. However, due to data privacy concerns, police stations do not disclose or provide open access to FIR data. This necessitates the development of methods that analyze FIR data while preserving privacy, leading to the adoption of Federated Learning (FL) as an effective solution. FL allows data to remain decentralized while enabling collaborative model training, thus addressing privacy challenges. In this study, historical FIR data from Karnataka State is utilized to train Machine Learning (ML) and Deep Learning (DL) models for predicting convictions. Various classification algorithms are implemented, including Logistic Regression, Decision Tree, Naïve Bayes, Random Forest, XGBoost, and Neural Networks. Random Forest achieved the highest classification performance, with an accuracy of 98% on test set, highlighting the model’s potential for deployment in privacy-sensitive environments. On implementing various FL methods on Random Forest, 95% was achieved with FedAvg. This study demonstrates the feasibility of privacy-preserving conviction prediction using decentralized FIR data.